This work was supported by the German Research Foundation (DFG, Collaborative Research Centre CRC 1181, subproject Z02; Priority Programme &#956;Bone, projects BA 4027/10-1 and BO 3811), including additional support for the scanning devices (INST&#160;410/77-1&#160;FUGG and INST&#160;410/93-1&#160;FUGG), and by the Emerging Fields Initiative (EFI) &#8220;Big Thera&#8221; of the Friedrich-Alexander-University Erlangen-N&#252;rnberg.

All care and experimental procedures were performed in accordance with national and regional legislation on animal protection, and all animal procedures were approved by the State Government of Franconia, Germany (reference number 55.2 DMS-2532-2-228).
1. Induction of breast cancer bone metastases in the right hind leg of nude rats
NOTE: A detailed description of the induction of breast cancer bone metastases in nude rats has been published elsewhere6,8. The most relevant steps are presented below.
<ol>
	<li>Culture MDA-MB-231 human breast cancer cells in RPMI-1640, supplemented with 10% fetal calf serum (FCS). Keep the cells under standard conditions (37 &#176;C, 5% CO2) and passage the cells 2&#8211;3 times a week.</li>
	<li>Wash near-confluent MDA-MB-231 cells with 2 mM EDTA in phosphate-buffered saline (PBS), and then detach the cells with 0.25% trypsin. Determine the cell concentration with a Neubauer&#8217;s chamber and resuspend them in 200 &#181;L of RPMI-1640 at a concentration of 1.5 x 105 cells/200&#160;&#181;L.</li>
	<li>Use 6&#8211;8 week-old nude rats and keep them under pathogen-free, controlled conditions (21 &#176;C &#177; 2 &#176;C room temperature, 60% humidity, and 12 h light-dark rhythm). Offer autoclaved feed and water ad libitum.</li>
	<li>Before performing the surgery, inject an analgesic drug (e.g., Carprofen 4 mg/kg) subcutaneously. Anesthetize rats with an isoflurane (1&#8211;1.5 vol. %)/oxygen mixture at a flow rate of 2 L/min. Check the anesthetic depth by toe pinching.</li>
	<li>For surgery, use an operating microscope with a 16x magnification.</li>
	<li>Perform a 2&#8211;3 cm cut in the rat&#8217;s right inguinal region. Dissect all arteries in the right inguinal region, including the femoral artery (FA), the superficial epigastric artery (SEA), the descending genicular artery (DGA), the popliteal artery (PA), and the saphenous artery (SA). Place two removable clips on the FA: one proximal to the beginning of the SEA, and another directly proximal to the beginning of the DGA.</li>
	<li>Ligate the distal portion of the SEA. Perform a cut of the SEA&#8217;s wall and insert a 0.3&#160;mm diameter needle into the SEA. Connect a syringe containing the cell suspension from step 1.2 to the needle. Remove the distal clip from the FA and clip the SA instead.</li>
	<li>Slowly inject the MDA-MB-231 cell suspension from step 1.2 (1.5 x 105 cells/200&#160;&#181;L) into the SEA. Remove the needle, ligate the SEA, and remove the artery clips. Close the wound using surgical clips and terminate anesthesia. Monitor the animals daily to assess tumor size and any evidence of pain.</li>
</ol>
2. Magnetic resonance imaging (MRI)
NOTE: For a detailed description of MRI procedures, please see B&#228;uerle et al.11.
<ol>
	<li>Perform MRI 10 days PI using a dedicated experimental scanner (see Table of Materials) or a human MR system with an appropriate animal coil.</li>
	<li>Anesthetize the rat with an isoflurane (1&#8211;1.5 vol. %)/oxygen mixture as described above. Place a catheter in the rat&#8217;s tail vein and tape it to the tail. Connect a syringe containing the contrast agent (0.1 mmol/kg Gd-DTPA in approximately 0.5 mL).</li>
	<li>Place the anesthetized rat in the MR system. Locate the distal femur and proximal tibia of the right hind leg in an anatomic sequence (e.g., T2-weighted turbo spin echo sequence; TR = 8,654 ms; TE = 37 ms; matrix 320 x 272; FOV = 65 mm x 55 mm; slice thickness = 1&#160;mm; scan time 11:24 min).</li>
	<li>Determine the slices covering the distal femur and proximal tibia of the right hind leg and start the DCE-MRI sequence (e.g., fast low angle shot sequence; TR = 3.9 ms; TE = 0.88 ms; matrix = 256 x 216; FOV = 65 x 54&#160;mm2; slice thickness = 1 mm; 8 slices; 100 time points; scan time = 8:25&#160;min). After 30 s, start injecting the contrast agent over a time period of 10&#160;s. 
	NOTE: The total time to perform an MRI examination is approximately 20 min per animal.</li>
</ol>
3. Positron emission tomography/computed tomography (PET/CT)
NOTE: For a detailed description of the PET procedures, please see Cheng at al.12.
<ol>
	<li>Perform PET/CT imaging 10 days PI using a dedicated experimental scanner (see Table of Materials).</li>
	<li>Keep the animals fasted prior to imaging. Anesthetize the rat as described in step 2.2 and insert a catheter in the tail vein as described above.</li>
	<li>Inject 6&#160;MBq of 18F-Fluorodeoxyglucose (18F-FDG) into the tail vein and wait ~30&#160;min to allow the tracer to distribute properly.</li>
	<li>Perform a CT acquisition (tube voltage = 80&#160;kV, tube current = 500&#160;&#181;A, isotropic resolution = 48.9&#160;&#181;m, duration = 10&#160;min).</li>
	<li>Perform a static PET acquisition (lower/upper discriminatory level = 350/650&#160;keV; timing window = 3.438&#160;ns; duration = 15&#160;min).</li>
</ol>
4. Alternative imaging strategies
<ol>
	<li>For an early assessment of MDA-MB-231 cells in the hind leg, inoculate 1.5 x 105 labeled cells /200&#160;&#181;L for bioluminescence (i.e., cells expressing luciferin, MDA-MB-231-LUC13) or fluorescence imaging (i.e., cells expressing green or red fluorescent protein, MDA-MB-231-GFP/RFP13). Use the system for preclinical optical imaging to detect intraosseous MDA-MB-231 cells after tumor cell inoculation14.&#160;</li>
	<li>Perform experimental ultrasound using a dedicated scanner after intravenous injection of microbubbles to derive morphological and functional parameters of vascularization comparable to MRI7.</li>
</ol>
5. MRI analysis
<ol>
	<li>Use a DICOM viewer15 with a DCE Plugin16 and load the DCE sequence in 4D-mode by clicking the &#8220;Import&#8221; button in the top menu, selecting the DICOM folder containing the MR images from step 2.4, and clicking &#8220;4D Viewer&#8221; in the top menu.</li>
	<li>Place a circular 2-dimensional region of interest (ROI), with a target size of 1.5 mm2, in the proximal tibial shaft&#8217;s bone marrow of the right hind leg, preferably using image numbers 4 or 5 from the sequence consisting of 8 images, as these center images provide more stable results.</li>
	<li>Start the DCE plugin from the top menu, select &#8220;Relative Enhancement&#8221; in the &#8220;Plot Type&#8221; field, and define the baseline range from time points 1 to 5 by typing these numbers into the respective fields. Export the analysis as a .txt file with the respective button and choose &#8220;DCEraw.txt&#8221; as the file name.</li>
	<li>Open RStudio17 and load the provided DCE-Script.R file via the &#8220;File&#8221; menu by selecting &#8220;Open File&#8221;. Run the entire script by selecting &#8220;Code&#8221;, then &#8220;Run Region&#8221; and then &#8220;Run All&#8221; from the menu. Copy the output to the provided template file named &#8220;ImagingFeatures.xlsx&#8221; (Figure 2).</li>
	<li>In the DICOM viewer, place a second ROI within the back muscle of the animal and repeat steps 5.2&#8211;5.4 to obtain the muscle DCE measurements for normalization purposes. Within the spreadsheet &#8220;ImagingFeatures.xlsx&#8221;, the respective bone measurements are automatically divided by the respective muscle measurements for normalization purposes.</li>
	<li>Repeat steps 5.1&#8211;5.5 for all animals and complete the spreadsheet.</li>
</ol>
6. PET/CT analysis
<ol>
	<li>Open the PET/CT analysis software and import the data obtained in step 3 by clicking &#8220;File&#8221;, followed by &#8220;Manual import&#8221;. Mark the ct.img.hd and the pet.img.hdr files. Click &#8220;Open&#8221; and select &#8220;Import all&#8221;.</li>
	<li>Open the datasets by selecting &#8220;General analysis&#8221;, followed by &#8220;OK&#8221;.</li>
	<li>Select &#8220;ROI Quantification&#8221;, followed by &#8220;Create&#8221;, and then &#8220;Create a ROI from a template&#8221;. Place a 2-dimensional ROI approximately 4 mm x 6&#160;mm into the proximal tibial shaft&#8217;s bone marrow of the right hind leg.</li>
	<li>Select &#8220;ROIs (Target 1 overlay)&#8221; and write down mean, minimum, and maximum values in Bq/mL.</li>
	<li>Calculate the maximum standardized uptake value (SUVmax): Divide the maximum value (Bq/mL) by the injected activity and multiply the result by the weight of the animal in grams. Enter the result into the spreadsheet (Figure 2).</li>
</ol>
7. Determining the tumor-take rate
<ol>
	<li>To diagnose tumor growth in the right hind leg, repeat MR and PET/CT imaging on day 30&#160;PI, as described above. 
	NOTE: Tumors will be clearly visible on day 30 PI and feature T2w-hyperintense lesions and clear contrast enhancement in MRI, along with a clearly elevated SUVmax in PET/CT. According to previous experiments, 60%&#8211;80% of the animals will develop metastases in their right hind leg.</li>
	<li>Complete the spreadsheet by adding an additional &#8220;Tumor&#8221; column and enter &#8220;1&#8221; for every animal that presents metastases, and &#8220;0&#8221; for every animal without visible tumor burden (Figure 2). Save the spreadsheet as &#8220;ImagingFeatures.xlsx&#8221; within the Downloads folder.</li>
</ol>
8. Feature selection
<ol>
	<li>To determine the most relevant features for prediction of future tumor growth, import the spreadsheet into an open-source data visualization, machine learning, and data mining toolkit18.</li>
	<li>Draw the File-subroutine from the Data menu into the workspace on the right and double-click it. Load the spreadsheet by clicking the &#8220;Folder&#8221; icon and selecting the file &#8220;ImagingFeatures.xlsx&#8221;. Select the &#8220;Export&#8221; worksheet and assign the target-attribute to the variable &#8220;Tumor&#8221;. Assign the &#8220;Skip&#8221; function to the animal number (Figure 3).</li>
	<li>Draw the &#8220;Rank&#8221; subroutine from the Data menu into the workspace and connect the &#8220;File&#8221; and &#8220;Rank&#8221; subroutines by drawing a line between them.</li>
	<li>Open the &#8220;Rank&#8221; subroutine by double-clicking on its icon, and select the &#8220;Information Gain&#8221; algorithm19.</li>
	<li>From the five acquired parameters, use the top three for further analyses (SUVmax, PE, and AUC). 
	Note: These parameters reflect metabolic activity (SUVmax) and tissue vascularization (PE and AUC).</li>
</ol>
9. ML analysis
<ol>
	<li>Open RStudio 3.4.117 and load the provided TrainModel.R-Script via the &#8220;File&#8221; menu.</li>
	<li>Install the required libraries (this only has to be done once) by typing:&#160;install.packages(c(&#34;caret&#34;, &#34;readxl&#34;, &#34;pROC&#34;, &#34;RcmdrPlugin.EZR&#34;, &#34;ggplot2&#34;))</li>
	<li>To load the required libraries and set the Downloads folder as the working directory, select the lines 3&#8211;5 within the TrainModel.R Script.</li>
	<li>Run the selected code by clicking &#8220;Code&#8221; within the menu, and then &#8220;Run Selected Line(s)&#8221;.</li>
</ol>
10. Training an avNNet ML algorithm
<ol>
	<li>To train an avNNet algorithm, select the lines 8&#8211;39 from the TrainModel.R-Script (see step 9.1).</li>
	<li>Run the selected code by clicking &#8220;Code&#8221; within the menu, and then &#8220;Run Selected Line(s)&#8221;.</li>
</ol>
11. Analyzing the ML algorithm&#8217;s results
<ol>
	<li>To assess standard parameters of diagnostic accuracy (sensitivity, specificity, positive and negative predictive values, and likelihood ratios), select the lines 41&#8211;50 from the TrainModel.R-Script.</li>
	<li>Run the selected code by clicking &#8220;Code&#8221; within the menu, and then &#8220;Run Selected Line(s)&#8221;.</li>
</ol>
12. Comparing the final model&#39;s Receiver Operating Characteristic (ROC) curve with the ROC curves of its constituent parameters
<ol>
	<li>To perform DeLong&#8217;s tests to compare the model&#8217;s ROC curve with the ROC curves of its constituent parameters, select the lines 52&#8211;62 from the TrainModel.R-Script (see step 9.1).</li>
	<li>Run the selected code by clicking &#8220;Code&#8221; within the menu, and then &#8220;Run Selected Line(s)&#8221;.</li>
</ol>

<ol><li>D'Oronzo, S., Brown, J., Coleman, R. <a target="_blank" href="http://www.ncbi.nlm.nih.gov/pubmed?term=((The+role+of+biomarkers+in+the+management+of+bone-homing+malignancies%5BTitle%5D)+AND+%22Journal+of+Bone+Oncology%22%5BJournal%5D)">The role of biomarkers in the management of bone-homing malignancies.</a> Journal of Bone Oncology. 9, 1-9 (2017).</li>
<li>Ellmann, S., Beck, M., Kuwert, T., Uder, M., Bäuerle, T. <a target="_blank" href="http://www.ncbi.nlm.nih.gov/pubmed?term=((Multimodal+imaging+of+bone+metastases%3A+From+preclinical+to+clinical+applications%5BTitle%5D)+AND+%22Journal+of+Orthopaedic+Translation%22%5BJournal%5D)">Multimodal imaging of bone metastases: From preclinical to clinical applications.</a> Journal of Orthopaedic Translation. 3 (4), 166-177 (2015).</li>
<li>Braun, S., Pantel, K. <a target="_blank" href="http://www.ncbi.nlm.nih.gov/pubmed?term=((Clinical+significance+of+occult+metastatic+cells+in+bone+marrow+of+breast+cancer+patients%5BTitle%5D)+AND+%22The+Oncologist%22%5BJournal%5D)">Clinical significance of occult metastatic cells in bone marrow of breast cancer patients.</a> The Oncologist. 6 (2), 125-132 (2001).</li>
<li>Braun, S., Rosenberg, R., Thorban, S., Harbeck, N. <a target="_blank" href="http://www.ncbi.nlm.nih.gov/pubmed?term=((Implications+of+occult+metastatic+cells+for+systemic+cancer+treatment+in+patients+with+breast+or+gastrointestinal+cancer%5BTitle%5D)+AND+%22Seminars+in+surgical+oncology%22%5BJournal%5D)">Implications of occult metastatic cells for systemic cancer treatment in patients with breast or gastrointestinal cancer.</a> Seminars in surgical oncology. 20 (4), 334-346 (2001).</li>
<li>Ellmann, S., et al. <a target="_blank" href="http://www.ncbi.nlm.nih.gov/pubmed?term=((Prediction+of+early+metastatic+disease+in+experimental+breast+cancer+bone+metastasis+by+combining+PET%2FCT+and+MRI+parameters+to+a+Model-Averaged+Neural+Network%5BTitle%5D)+AND+%22Bone%22%5BJournal%5D)">Prediction of early metastatic disease in experimental breast cancer bone metastasis by combining PET/CT and MRI parameters to a Model-Averaged Neural Network.</a> Bone. 120, 254-261 (2018).</li>
<li>Bäuerle, T., Komljenovic, D., Berger, M. R., Semmler, W. <a target="_blank" href="http://www.ncbi.nlm.nih.gov/pubmed?term=((Multi-modal+imaging+of+angiogenesis+in+a+nude+rat+model+of+breast+cancer+bone+metastasis+using+magnetic+resonance+imaging%2C+volumetric+computed+tomography+and+ultrasound%5BTitle%5D)+AND+%22Journal+of+Visualized+Experiments%22%5BJournal%5D)">Multi-modal imaging of angiogenesis in a nude rat model of breast cancer bone metastasis using magnetic resonance imaging, volumetric computed tomography and ultrasound.</a> Journal of Visualized Experiments. (66), e4178(2012).</li>
<li>Merz, M., Komljenovic, D., Semmler, W., Bäuerle, T. <a target="_blank" href="http://www.ncbi.nlm.nih.gov/pubmed?term=((Quantitative+contrast-enhanced+ultrasound+for+imaging+antiangiogenic+treatment+response+in+experimental+osteolytic+breast+cancer+bone+metastases%5BTitle%5D)+AND+%22Investigative+Radiology%22%5BJournal%5D)">Quantitative contrast-enhanced ultrasound for imaging antiangiogenic treatment response in experimental osteolytic breast cancer bone metastases.</a> Investigative Radiology. 47 (7), 422-429 (2012).</li>
<li>Bäuerle, T., et al. <a target="_blank" href="http://www.ncbi.nlm.nih.gov/pubmed?term=((Characterization+of+a+rat+model+with+site-specific+bone+metastasis+induced+by+MDA-MB-231+breast+cancer+cells+and+its+application+to+the+effects+of+an+antibody+against+bone+sialoprotein%5BTitle%5D)+AND+%22International+Journal+of+Cancer%22%5BJournal%5D)">Characterization of a rat model with site-specific bone metastasis induced by MDA-MB-231 breast cancer cells and its application to the effects of an antibody against bone sialoprotein.</a> International Journal of Cancer. 115 (2), 177-186 (2005).</li>
<li>Patel, J., Goyal, R. <a target="_blank" href="http://www.ncbi.nlm.nih.gov/pubmed?term=((Applications+of+Artificial+Neural+Networks+in+Medical+Science%5BTitle%5D)+AND+%22Current+Clinical+Pharmacology%22%5BJournal%5D)">Applications of Artificial Neural Networks in Medical Science.</a> Current Clinical Pharmacology. 2 (3), 217-226 (2008).</li>
<li>Naftaly, U., Intrator, N., Horn, D. <a target="_blank" href="http://www.ncbi.nlm.nih.gov/pubmed?term=((Optimal+ensemble+averaging+of+neural+networks%5BTitle%5D)+AND+%22Network%3A+Computation+in+Neural+Systems%22%5BJournal%5D)">Optimal ensemble averaging of neural networks.</a> Network: Computation in Neural Systems. 8 (3), 283-296 (1997).</li>
<li>Bäuerle, T., Merz, M., Komljenovic, D., Zwick, S., Semmler, W. <a target="_blank" href="http://www.ncbi.nlm.nih.gov/pubmed?term=((Drug-induced+vessel+remodeling+in+bone+metastases+as+assessed+by+dynamic+contrast+enhanced+magnetic+resonance+imaging+and+vessel+size+imaging%3A+A+longitudinal+in+vivo+study%5BTitle%5D)+AND+%22Clinical+Cancer+Research%22%5BJournal%5D)">Drug-induced vessel remodeling in bone metastases as assessed by dynamic contrast enhanced magnetic resonance imaging and vessel size imaging: A longitudinal in vivo study.</a> Clinical Cancer Research. 16 (12), 3215-3225 (2010).</li>
<li>Cheng, C., et al. <a target="_blank" href="http://www.ncbi.nlm.nih.gov/pubmed?term=((Evaluation+of+treatment+response+of+cilengitide+in+an+experimental+model+of+breast+cancer+bone+metastasis+using+dynamic+PET+with+18F-FDG%5BTitle%5D)+AND+%22Hellenic+Journal+of+Nuclear+Medicine%22%5BJournal%5D)">Evaluation of treatment response of cilengitide in an experimental model of breast cancer bone metastasis using dynamic PET with 18F-FDG.</a> Hellenic Journal of Nuclear Medicine. 14 (1), 15-20 (2011).</li>
<li>Marturano-Kruik, A., et al. <a target="_blank" href="http://www.ncbi.nlm.nih.gov/pubmed?term=((Human+bone+perivascular+niche-on-a-chip+for+studying+metastatic+colonization%5BTitle%5D)+AND+%22Proceedings+of+the+National+Academy+of+Sciences+of+the+United+States+of+America%22%5BJournal%5D)">Human bone perivascular niche-on-a-chip for studying metastatic colonization.</a> Proceedings of the National Academy of Sciences of the United States of America. 115 (6), 1256-1261 (2018).</li>
<li>Sonntag, E., et al. <a target="_blank" href="http://www.ncbi.nlm.nih.gov/pubmed?term=((In+vivo+proof-of-concept+for+two+experimental+antiviral+drugs%2C+both+directed+to+cellular+targets%2C+using+a+murine+cytomegalovirus+model%5BTitle%5D)+AND+%22Antiviral+Research%22%5BJournal%5D)">In vivo proof-of-concept for two experimental antiviral drugs, both directed to cellular targets, using a murine cytomegalovirus model.</a> Antiviral Research. 161, 63-69 (2019).</li>
<li>Horos - Free DICOM Medical Image Viewer | Open-Source. , Available from: 
 <a target="_blank" href="https://www.horosproject.org/">https://www.horosproject.org/</a> (2015).</li>
<li>DCE Tool Plugin. , Available from: 
 <a target="_blank" href="http://kyungs.bol.ucla.edu/software/DCE_tool/DCE_tool.html">http://kyungs.bol.ucla.edu/software/DCE_tool/DCE_tool.html</a> (2020).</li>
<li>RStudio Team RStudio: Inteegrated Development for R. , Available from: 
 <a target="_blank" href="https://rstudio.com">https://rstudio.com</a> (2015).</li>
<li>Demšar, J., et al. <a target="_blank" href="http://www.ncbi.nlm.nih.gov/pubmed?term=((Orange%3A+Data+Mining+Toolbox+in+Python%5BTitle%5D)+AND+%22Journal+of+Machine+Learning+Research%22%5BJournal%5D)">Orange: Data Mining Toolbox in Python.</a> Journal of Machine Learning Research. 14, 2349-2353 (2013).</li>
<li>Saeys, Y., Inza, I., Larrañaga, P. <a target="_blank" href="http://www.ncbi.nlm.nih.gov/pubmed?term=((A+review+of+feature+selection+techniques+in+bioinformatics%5BTitle%5D)+AND+%22Bioinformatics%22%5BJournal%5D)">A review of feature selection techniques in bioinformatics.</a> Bioinformatics. 23 (19), 2507-2517 (2007).</li>
<li>Kuhn, M. CRAN - Package caret. , Available from: 
 <a target="_blank" href="https://cran.r-project.org/web/packages/caret/index.html">https://cran.r-project.org/web/packages/caret/index.html</a> (2016).</li>
<li>Chen, T., et al. CRAN: Package xgboost - Extreme Gradient Boosting. , Available from: 
 <a target="_blank" href="https://cran.r-project.org/web/packages/xgboost/">https://cran.r-project.org/web/packages/xgboost/</a> (2019).</li>
<li>Available Models | The caret Package. , Available from: 
 <a target="_blank" href="http://topepo.github.io/caret/available-models.html">http://topepo.github.io/caret/available-models.html</a> (2020).</li>
<li>Fernández-Delgado, M., Cernadas, E., Barro, S., Amorim, D., Fernández-Delgado, A. <a target="_blank" href="http://www.ncbi.nlm.nih.gov/pubmed?term=((Do+we+Need+Hundreds+of+Classifiers+to+Solve+Real+World+Classification+Problems%5BTitle%5D)+AND+%22Journal+of+Machine+Learning+Research%22%5BJournal%5D)">Do we Need Hundreds of Classifiers to Solve Real World Classification Problems.</a> Journal of Machine Learning Research. 15, 3133-3181 (2014).</li>
<li>Hira, Z. M., Gillies, D. F. <a target="_blank" href="http://www.ncbi.nlm.nih.gov/pubmed?term=((A+Review+of+Feature+Selection+and+Feature+Extraction+Methods+Applied+on+Microarray+Data%5BTitle%5D)+AND+%22Advances+in+Bioinformatics%22%5BJournal%5D)">A Review of Feature Selection and Feature Extraction Methods Applied on Microarray Data.</a> Advances in Bioinformatics. 2015, 198363(2015).</li>
<li>Sánchez-Maroño, N., Alonso-Betanzos, A., Tombilla-Sanromán, M. <a target="_blank" href="http://www.ncbi.nlm.nih.gov/pubmed?term=((Filter+methods+for+feature+selection+-+A+comparative+study%5BTitle%5D)+AND+%22Lecture+Notes+in+Computer+Science+%28including+subseries+Lecture+Notes+in+Artificial+Intelligence+and+Lecture+Notes+in+Bioinformatics%29%22%5BJournal%5D)">Filter methods for feature selection - A comparative study.</a> Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics). 4881, 178-187 (2007).</li>
<li>Cawley, G. C., Talbot, N. L. C. C. <a target="_blank" href="http://www.ncbi.nlm.nih.gov/pubmed?term=((Fast+exact+leave-one-out+cross-validation+of+sparse+least-squares+support+vector+machines%5BTitle%5D)+AND+%22Neural+Network%22%5BJournal%5D)">Fast exact leave-one-out cross-validation of sparse least-squares support vector machines.</a> Neural Network. 17 (10), 1467-1475 (2004).</li>
<li>Forghani, R., et al. <a target="_blank" href="http://www.ncbi.nlm.nih.gov/pubmed?term=((Radiomics+and+Artificial+Intelligence+for+Biomarker+and+Prediction+Model+Development+in+Oncology%5BTitle%5D)+AND+%22Computational+and+Structural+Biotechnology+Journal%22%5BJournal%5D)">Radiomics and Artificial Intelligence for Biomarker and Prediction Model Development in Oncology.</a> Computational and Structural Biotechnology Journal. 17, 995-1008 (2019).</li>
<li>Jaffe, C. C. <a target="_blank" href="http://www.ncbi.nlm.nih.gov/pubmed?term=((Measures+of+response%3A+RECIST%2C+WHO%2C+and+new+alternatives%5BTitle%5D)+AND+%22Journal+of+Clinical+Oncology+Official+Journal+of+the+American+Society+of+Clinical+Oncology%22%5BJournal%5D)">Measures of response: RECIST, WHO, and new alternatives.</a> Journal of Clinical Oncology Official Journal of the American Society of Clinical Oncology. 24 (20), 3245-3251 (2006).</li>
<li>Lambin, P., et al. <a target="_blank" href="http://www.ncbi.nlm.nih.gov/pubmed?term=((Radiomics%3A+Extracting+more+information+from+medical+images+using+advanced+feature+analysis%5BTitle%5D)+AND+%22European+Journal+of+Cancer%22%5BJournal%5D)">Radiomics: Extracting more information from medical images using advanced feature analysis.</a> European Journal of Cancer. 48 (4), 441-446 (2012).</li>
<li>Gillies, R. J., Kinahan, P. E., Hricak, H. <a target="_blank" href="http://www.ncbi.nlm.nih.gov/pubmed?term=((Radiomics%3A+Images+are+more+than+pictures%2C+they+are+data%5BTitle%5D)+AND+%22Radiology%22%5BJournal%5D)">Radiomics: Images are more than pictures, they are data.</a> Radiology. 278 (2), 563-577 (2016).</li>
<li>Nioche, C., et al. <a target="_blank" href="http://www.ncbi.nlm.nih.gov/pubmed?term=((Lifex%3A+A+freeware+for+radiomic+feature+calculation+in+multimodality+imaging+to+accelerate+advances+in+the+characterization+of+tumor+heterogeneity%5BTitle%5D)+AND+%22Cancer+Research%22%5BJournal%5D)">Lifex: A freeware for radiomic feature calculation in multimodality imaging to accelerate advances in the characterization of tumor heterogeneity.</a> Cancer Research. 78 (16), 4786-4789 (2018).</li>
<li>Ellmann, S., et al. <a target="_blank" href="http://www.ncbi.nlm.nih.gov/pubmed?term=((Application+of+machine+learning+algorithms+for+multiparametric+MRI-based+evaluation+of+murine+colitis%5BTitle%5D)+AND+%22PLOS+ONE%22%5BJournal%5D)">Application of machine learning algorithms for multiparametric MRI-based evaluation of murine colitis.</a> PLOS ONE. 13 (10), 0206576(2018).</li>
</ol>

The authors declare no conflict of interest. The funders had no role in the design of the study; in the collection, analyses, or interpretation of data; in the writing of the manuscript, or in the decision to publish the results.

ML algorithms are powerful tools used to integrate several predictive features into a combined model and obtain an accuracy that exceeds that of its separate constituents when used alone. Nonetheless, the actual result depends on several critical steps. First, the ML algorithm used is a crucial factor, because different ML algorithms yield different results. The algorithm used in this protocol is an avNNet, but other promising algorithms include Extreme Gradient Boosting21 or Random Forests. The caret package20 for RStudio provides a plethora of different algorithms (currently &#62;175), and the proposed protocol is highly flexible in terms of switching from one algorithm to another by simply changing single lines of code (e.g., changing method = &#8220;avNNet&#8221; to method =&#8220;rf&#8221;) and adapting the TunedGrid-settings to the respective ML algorithm. For details, see the caret github repository22. An excellent overview of different algorithms and their performance with respect to different classification problems was published by Fern&#225;ndez-Delgado et al.23 and could serve as a starting point for other experiments.
Another crucial factor is the choice of relevant features to include in an ML algorithm. This protocol proposes the use of the filter method, &#8220;Information Gain&#8221;19, to rank the available features in descending order and use the most relevant ones to train the avNNet. Filter methods are based only on general assumptions, such as correlation with the variable to predict, so researchers should preselect features independently of the classifier used24,25. Such methods are particularly effective in computation time and robust to overfitting. However, the cutoff that separates relevant from irrelevant features is defined by the user, making it somewhat arbitrary. For the proposed protocol, the features with the top 75% information gain were used, corresponding to SUVmax, PE, and AUC. This selection, however, can be systematically strengthened by including a random variable that has no relationship to the target, calculating its information gain, and then comparing it to the information gain of the real features (Supplementary Figure 2). This slightly more sophisticated approach additionally confirmed the choice of the three abovementioned features as being the most relevant. Nonetheless, several different filter methods exist, along with other approaches that select features with respect to a particular classifier algorithm, such as feature extraction and wrapper methods. Different feature selection approaches may yield different results.
To ensure generalizability of the ML algorithm and further prevent overfitting, the proposed protocol applies leave-one-out cross-validation (LOOCV). The best approach, however, would be to randomly remove a subset from the entire data set, and treat it as a testing set. The ML algorithm is then trained on the remainder of the data (i.e., the training set) to subsequently predict the outcome of the testing set. However, this approach needs a sufficiently large data set. For smaller sample sizes, application of LOOCV is common because it provides an almost unbiased estimate of a model&#39;s true generalization ability26. In LOOCV, the first data point is removed from the data set, and the avNNet is trained with the retained data. Then, the outcome of the formerly withheld data point is predicted and saved. This process is repeated for all data points, so that finally each outcome is predicted with data that were not used for training the algorithm. Other validation approaches include x-fold cross-validations (most commonly 10-fold) and can be easily applied by changing the respective trainControl parameter within the code to method=&#34;CV&#34;.
From a quantitative point-of-view, medical images are typically evaluated in a very basic way, largely relying on measurements of size and shape of potentially suspicious lesions or areas27,28. However, the advantage of the Digital Imaging and Communications in Medicine (DICOM) standard is that it allows extraction of many features, referred to as radiomics. The term &#8220;radiomics&#8220; was initially defined as the high-throughput extraction of large quantities of image features29, but was subsequently extended to include the conversion of images to higher dimensional data30. However, the higher dimensional data are mainly used to identify correlations rather than causes30. The features described in this protocol fall in between classic radiological features, such as size and shape, and radiomics, as they resemble generally accepted parameters of vascularization and metabolic activity. This offers a potential causal relation to the microinvasion of disseminated tumor cells. If desired by the user, an extraction of radiomic features can be performed with different software packages31.
The protocol provided is not restricted to a finite number of features. Thus, it can be used with large radiomics data sets. However, the abovementioned issue of feature selection becomes increasingly important with growing data sets. The presented protocol can also be transferred to different study settings, e.g., from the fields of oncology, infection, or inflammation32.

The purpose of this protocol is to integrate several functional imaging parameters from MRI and PET/CT into a model-averaged neural network (avNNet) ML algorithm. This algorithm predicts the growth of macrometastases in a rat model of breast cancer bone metastases at an early timepoint, when macroscopic changes within the bone are not yet visible.
Prior to the growth of macrometastases, a bone marrow invasion of disseminated tumor cells occurs, commonly referred to as micrometastatic disease1,2. This initial invasion can be considered an early step in metastatic disease, but is typically missed during conventional staging examinations3,4. Although the currently available imaging modalities cannot detect bone marrow microinvasion when used alone, a combination of imaging parameters yielding information on vascularization and metabolic activity has been shown to perform better5. This complementary benefit is achieved by combining different imaging parameters into an avNNet, which is an ML algorithm. Such an avNNet allows for the reliable prediction of bone macrometastases formation before any visible metastases are present. Therefore, integrating imaging biomarkers into an avNNet could serve as a surrogate parameter for bone marrow microinvasion and early metastatic disease.
To develop the protocol, a previously described model of breast cancer bone metastases in nude rats was used6,7,8. The advantage of this model is its site-specificity, meaning that the animals develop bony metastases exclusively in their right hind leg. However, the tumor-take rate of this approach is 60%&#8211;80%, so a considerable number of the animals do not develop any metastases during the study. Using imaging modalities such as MRI and PET/CT, the presence of metastases is detectable from day 30 postinjection (PI). At earlier time points (e.g., 10 PI) imaging does not distinguish between animals that will develop metastatic disease and those will not (Figure 1).
An avNNet trained on functional imaging parameters acquired on day 10 PI, as described in the following protocol, reliably predicts or excludes the growth of macrometastases within the following ~3 weeks. Neural Networks combine artificial nodes within different layers. In the study protocol, the functional imaging parameters for bone marrow blood supply and metabolic activity represent the bottom layer, while the prediction of malignancy represents the top layer. An additional intermediate layer contains hidden nodes that are connected to both the top and the bottom layer. The strength of the connections between the different nodes is updated during the training of the network to perform the respective classification task with high accuracy9. The accuracy of such a neural network can be further increased by averaging the outputs of several models, resulting in an avNNet10.

<table><tbody><tr><td>Binocular Operating Microscope</td><td>Leica</td><td>NA</td><td></td></tr><tr><td>ClinScan MR System</td><td>Bruker</td><td>NA</td><td></td></tr><tr><td>DICOM Viewer</td><td>Horos</td><td>NA</td><td>www.horosproject.org</td></tr><tr><td>Excel: Spreadsheet</td><td>Microsoft</td><td>NA</td><td></td></tr><tr><td>FCS</td><td>Sigma</td><td>F2442-500ML</td><td></td></tr><tr><td>Gadovist</td><td>Bayer-Schering</td><td>NA</td><td></td></tr><tr><td>Inveon PET/CT</td><td>Siemens</td><td>NA</td><td></td></tr><tr><td>Inveon Research Workplace Software</td><td>Siemens Healthcare GmbH</td><td>NA</td><td></td></tr><tr><td>IVIS Spectrum</td><td>PerkinElmer</td><td>NA</td><td></td></tr><tr><td>MDA-MB-231 human breast cancer cells</td><td>American Type Culture Collection</td><td>N/A</td><td></td></tr><tr><td>Open-source data visualization, machine learning and data mining toolkit.</td><td>Orange3, University of Ljubljana</td><td>NA</td><td>https://orange.biolab.si/</td></tr><tr><td>RPMI-1640</td><td>Invitrogen/ThermoFisher</td><td>11875093</td><td></td></tr><tr><td>Trypsin</td><td>Sigma</td><td>9002-07-7</td><td></td></tr><tr><td>Vevo 3100</td><td>VisualSonics</td><td>NA</td><td></td></tr></tbody></table>

machine learning algorithms for early detection of bone metastases in an experimental rat model

Machine learning (ML) algorithms permit the integration of different features into a model to perform classification or regression tasks with an accuracy exceeding its constituents. This protocol describes the development of an ML algorithm to predict the growth of breast cancer bone macrometastases in a rat model before any abnormalities are observable with standard imaging methods. Such an algorithm can facilitate the detection of early metastatic disease (i.e., micrometastasis) that is regularly missed during staging examinations.
The applied metastasis model is site-specific, meaning that the rats develop metastases exclusively in their right hind leg. The model&#8217;s tumor-take rate is 60%&#8211;80%, with macrometastases becoming visible in magnetic resonance imaging (MRI) and positron emission tomography/computed tomography (PET/CT) in a subset of animals 30 days after induction, whereas a second subset of animals exhibit no tumor growth.
Starting from image examinations acquired at an earlier time point, this protocol describes the extraction of features that indicate tissue vascularization detected by MRI, glucose metabolism by PET/CT, and the subsequent determination of the most relevant features for the prediction of macrometastatic disease. These features are then fed into a model-averaged neural network (avNNet) to classify the animals into one of two groups: one that will develop metastases and the other that will not develop any tumors. The protocol also describes the calculation of standard diagnostic parameters, such as overall accuracy, sensitivity, specificity, negative/positive predictive values, likelihood ratios, and the development of a receiver operating characteristic. An advantage of the proposed protocol is its flexibility, as it can be easily adapted to train a plethora of different ML algorithms with adjustable combinations of an unlimited number of features. Moreover, it can be used to analyze different problems in oncology, infection, and inflammation.

The rats recovered quickly from the surgery and injection of the MDA-MB-231 breast cancer cells and were then subjected to MR- and PET/CT imaging on days 10 and 30 PI (Figure 1). A representative DCE analysis of a rat&#8217;s right proximal tibia is presented in Figure 2A. The DCE raw measurements were saved by selecting the &#8220;Export&#8221; button and choosing &#8220;DCEraw.txt&#8221; as the file name.
Subsequent calculations of the dynamic parameters, AUC, PE, and washout were performed in RStudio with the respective script. The output of the DCE measurements had to be saved as &#8220;DCEraw.txt&#8221; within the &#8220;Downloads&#8221; folder, so that the script could be run directly without additional configurations to provide a data table, as depicted in Figure 2B. These data were copied to the provided spreadsheet (Figure 2C). Similarly, the DCE parameters for muscular tissue were determined and transferred to the spreadsheet (Figure 2D,E). These values were normalized by dividing the bone measurements by the muscle measurements; this was performed automatically within the spreadsheet. From the PET/CT, the calculated SUVmax values were subsequently transferred into the table (Figure 2F).
On day 30 PI, all animals were evaluated to determine whether or not they developed metastases, and the table was completed by coding positive tumor burden as &#8220;1&#8221; and healthy animals as &#8220;0&#8221; within the rightmost column of the spreadsheet (Figure 2C).
The spreadsheet was imported into the open-source data visualization, machine learning, and data mining toolkit, and feature ranking revealed the SUVmax, PE, and AUC as the top three features for prediction of metastatic disease (Figure 3). These parameters reflect metabolic activity (SUVmax) and tissue vascularization (PE and AUC).
Running the TrainModel.R Script automatically imported the spreadsheet and calculated an avNNet. The optimal hyperparameter combination was determined (Figure 4A) and the final model was then calculated using the optimal hyperparameter combination (Figure 4B). Subsequently, a set of standard diagnostic parameters was calculated (Figure 4C) and an ROC curve of the model was plotted (Figure 4D).
The positive result is shown in Figure 4B&#8211;D. A comparison of the model&#8217;s ROC curve with the ROC curve of its three constituents (i.e., AUC, PE, and SUVmax) revealed that the model performed significantly better than all of its three constituents (p = 0.01 for AUC, p = 0.003 for PE, and p = 0.007 for SUVmax). The combination of the three selected parameters to an avNNet was more sensitive, thus allowing prediction of macroscopic disease with an overall accuracy of 85.7% (95% confidence interval = 67.3%&#8211;96.0%). These results were obtained from an analysis of 28 samples. The confidence intervals can be further narrowed by increasing the number of animals.
The negative results could be obtained as described here. The accuracy measures were highly sensitive to specific types of machine learning algorithms and to steps of data preprocessing. Neural Networks, in particular, tended to perform better when the input data were normalized. This was achieved by the &#8220;BoxCox&#8221; function in section 10 of the protocol (lines 22 and 36 within the provided TrainModel.R-Script). Refraining from normalization and using a different algorithm (e.g., a neural network not averaged), by changing the method to &#8220;nnet&#8221; (lines 21 and 35 within the provided TrainModel.R-Script), resulted in an area of 0.594 under the ROC curve (Supplementary Figure 1). Such a model failed to outperform its three constituents significantly (all p &#62; 0.15).
Because the script was optimized for RStudio 3.4.1 and the caret package version 6.0-84, using different software versions might yield different results. Running the provided scripts with the software versions used in this manuscript will give similar results. However, if readers aim to modify the script, add additional variables, change document folders or file names, or modify the machine learning algorithms in greater detail, respective adjustments of the code will be necessary. For these cases, the manual of the caret-package offers in-depth explanations20.
<img alt="Figure 1" class="xfigimg" src="/files/ftp_upload/61235/61235fig1.jpg" /> 
Figure 1: Representative MR and PET/CT images. MR and PET/CT images of the right hind leg of an animal that did not develop metastases over the course of the study (two leftmost columns, with images from day 10 and day 30 PI), and an animal that developed metastases between day 10 and day 30 PI (two rightmost columns, metastases marked with arrows). Note the high signal intensity of metastases in the T2w images (upper row), the contrast enhancement depicted by the increased area under the curve (AUC; second row), and the increased maximum standardized uptake value in the PET/CT (SUVmax; third row). Note that there are no visible differences in the images acquired on day 10 PI (first and third column) between the animal with metastases on day 30 PI and the animal that developed no bone metastases. This figure was modified from Ellmann et al.5. <a href="https://www-jove-com.remotexs.ntu.edu.sg/files/ftp_upload/61235/61235fig1large.jpg" target="_blank">Please click here to view a larger version of this figure.</a>
<img alt="Figure 2" class="xfigimg" src="/files/ftp_upload/61235/61235fig2.jpg" /> 
Figure 2: Assessment of the imaging features and compilation into a spreadsheet. (A) The dynamic contrast enhancement of the proximal tibia&#8217;s bone marrow was analyzed with a freeware DICOM viewer15 using a DCE plugin16. The respective measurements were saved, and (B) further analyzed with the provided DCE-Script.R-file in RStudio17. (C) The output was copied into a spreadsheet (see supplementary material for a template). (D) Likewise, the DCE measurement was performed for adjacent muscular tissue, analyzed using RStudio (E), and then copied to the spreadsheet. Data were normalized by dividing the results of the bone measurements by the results of the muscle measurements (C; salmon-shaded cells). (F) The PET/CT measurements were performed with the vendor&#8217;s software. The maximum standardized uptake value was calculated by dividing the respective measurement by the injected activity and multiplying it by the animal&#8217;s body weight. The result was subsequently copied into the spreadsheet. <a href="https://www-jove-com.remotexs.ntu.edu.sg/files/ftp_upload/61235/61235fig2large.jpg" target="_blank">Please click here to view a larger version of this figure.</a>
<img alt="Figure 3" class="xfigimg" src="/files/ftp_upload/61235/61235fig3.jpg" /> 
Figure 3: Feature ranking. Ranking of the acquired imaging features was performed within an open-source data visualization, machine learning, and data mining toolkit18 by importing the spreadsheet via the &#8220;File&#8221;-subroutine, and analyzing it via the &#8220;Rank&#8221;-subroutine. <a href="https://www-jove-com.remotexs.ntu.edu.sg/files/ftp_upload/61235/61235fig3large.jpg" target="_blank">Please click here to view a larger version of this figure.</a>
<img alt="Figure 4" class="xfigimg" src="/files/ftp_upload/61235/61235fig4.jpg" /> 
Figure 4: Representative RStudio output. The machine learning algorithm was developed using RStudio17 with the provided TrainModel.R-Script file. (A) A grid search among different hyperparameter combinations for the model-averaged neural network revealed a size of three neurons and a decay of 0.0005 as an optimum. (B) Using this hyperparameter combination, a full network was trained and cross-validated, reaching an overall accuracy of 85.7%. (C) Standard parameters of diagnostic accuracy, including sensitivity, specificity, positive and negative predictive values, and likelihood ratios, were calculated from a confusion matrix. (D) A representative ROC plot of the cross-validated model revealed an area under the curve (AUC) of 0.917. <a href="https://www-jove-com.remotexs.ntu.edu.sg/files/ftp_upload/61235/61235fig4large.jpg" target="_blank">Please click here to view a larger version of this figure.</a>
Supplementary Figure 1: Negative result. Changing the ML algorithm to a Neural Network without averaging and refraining from normalization of the input parameters led to a drop of the area under the curve of the ROC curve from 0.917 (Figure 4D) to 0.594. <a href="https://www-jove-com.remotexs.ntu.edu.sg/files/ftp_upload/61235/61235supfig1large.jpg" target="_blank">Please click here to download this file.</a>
Supplementary Figure 2: Alternative feature ranking. In contrast to the standard method depicted in Figure 3, a random variable was also introduced (&#8220;Random&#8221;; highlighted in blue), with its importance included in the ranking. This approach confirmed the applied selection of the variables SUVmax, PE, and AUC. <a href="https://www-jove-com.remotexs.ntu.edu.sg/files/ftp_upload/61235/61235supfig2large.jpg" target="_blank">Please click here to download this file.</a>

Watch this Scientific Journal Video about Machine Learning Algorithms for Early Detection of Bone Metastases in an Experimental Rat Model at JoVE.com

Machine Learning Algorithms for Early Detection of Bone Metastases in an Experimental Rat Model

This protocol was designed to train a machine learning algorithm to use a combination of imaging parameters derived from magnetic resonance imaging (MRI) and positron emission tomography/computed tomography (PET/CT) in a rat model of breast cancer bone metastases to detect early metastatic disease and predict subsequent progression to macrometastases.

Machine learning (ML) algorithms permit the integration of different features into a model to perform classification or regression tasks with an accuracy ...

machine-learning-algorithms-for-early-detection-bone-metastases-an

Nanyang Technological University

Research

JoVE Journal

Cancer Research

6.7K Views.  Friedrich-Alexander Universität Erlangen-Nürnberg. Dit protocol vergemakkelijkt de ontwikkeling van een machine learning-algoritme voor het voorspellen van de groei van botmetastasen in een experimenteel model in het stadium van vroege orgaankolonisatie. Het belangrijkste voordeel van deze techniek is dat het verschillende beeldvormingsparameters combineert in een machine learning-algoritme dat aanzienlijk beter presteert dan het productieve vermogen van elke afzonderlijke parameter. Hoewel dit protocol gericht is op de vroege diagnose van me...

Video: Machine Learning Algorithms for Early Detection of Bone Metastases in an Experimental Rat Model

Intra-iliac Artery Injection for Efficient and Selective Modeling of Microscopic Bone Metastasis

Intra-iliac artery (IIA) injection is an efficient approach to introduce metastatic lesions of various cancer cells in animals. Compared to the widely used intra-cardiac and intra-tibial injections, IIA injection brings several advantages. First, it can deliver a large quantity of cancer cells specifically to hind limb bones, thereby providing spatiotemporally synchronized early-stage colonization events and allowing robust quantification and swift detection of disseminated tumor cells. Second, it injects cancer cells into the circulation without damaging the local tissues, thereby avoiding inflammatory and wound-healing processes that confound the bone colonization process. Third, IIA injection causes very little metastatic growth in non-bone organs, thereby preventing animals from succumbing to other vital metastases, and allowing continuous monitoring of indolent bone lesions. These advantages are especially useful for the inspection of progression from single cancer cells to multi-cell micrometastases, which has largely been elusive in the past. When combined with cutting-edge approaches of biological imaging and bone histology, IIA injection can be applied to various research purposes related to bone metastases.

This manuscript provides the detailed procedure of intra-iliac artery (IIA) injection, a technique to deliver cancer cells specifically to hind limb tissues including bones to establish experimental bone metastases. Although initially established with breast tumor models, this protocol can be easily extended to other cancer types.

Intra-iliac artery (IIA) injection is an efficient approach to introduce metastatic lesions of various cancer cells in animals. Compared to the widely ...

In Vivo Model for Testing Effect of Hypoxia on Tumor Metastasis

Hypoxia has been implicated in the metastasis of Ewing sarcoma (ES) by clinical observations and in vitro data, yet direct evidence for its pro-metastatic effect is lacking and the exact mechanisms of its action are unclear. Here, we report an animal model that allows for direct testing of the effects of tumor hypoxia on ES dissemination and investigation into the underlying pathways involved. This approach combines two well-established experimental strategies, orthotopic xenografting of ES cells and femoral artery ligation (FAL), which induces hindlimb ischemia. Human ES cells were injected into the gastrocnemius muscles of SCID/beige mice and the primary tumors were allowed to grow to a size of 250 mm3. At this stage either the tumors were excised (control group) or the animals were subjected to FAL to create tumor hypoxia, followed by tumor excision 3 days later. The efficiency of FAL was confirmed by a significant increase in binding of hypoxyprobe-1 in the tumor tissue, severe tumor necrosis and complete inhibition of primary tumor growth. Importantly, despite these direct effects of ischemia, an enhanced dissemination of tumor cells from the hypoxic tumors was observed. This experimental strategy enables comparative analysis of the metastatic properties of primary tumors of the same size, yet significantly different levels of hypoxia. It also provides a new platform to further assess the mechanistic basis for the hypoxia-induced alterations that occur during metastatic tumor progression in vivo. In addition, while this model was established using ES cells, we anticipate that this experimental strategy can be used to test the effect of hypoxia in other sarcomas, as well as tumors orthotopically implanted in sites with a well-defined blood supply route.

In Vivo Model for Testing Effect of Hypoxia on Tumor Metastasis

This manuscript describes the development of an animal model that allows for the direct testing of the effects of tumor hypoxia on metastasis and the deciphering the mechanisms of its action. Although the experiments described here focus on Ewing sarcoma, a similar approach can be applied to other tumor types.

Hypoxia has been implicated in the metastasis of Ewing sarcoma (ES) by clinical observations and in vitro data, yet direct evidence for its pro-metastatic ...

Multi-modal Imaging of Angiogenesis in a Nude Rat Model of Breast Cancer Bone Metastasis Using Magnetic Resonance Imaging, Volumetric Computed Tomography and Ultrasound

Angiogenesis is an essential feature of cancer growth and metastasis formation. In bone metastasis, angiogenic factors are pivotal for tumor cell proliferation in the bone marrow cavity as well as for interaction of tumor and bone cells resulting in local bone destruction. Our aim was to develop a model of experimental bone metastasis that allows in vivo assessment of angiogenesis in skeletal lesions using non-invasive imaging techniques.
For this purpose, we injected 105 MDA-MB-231 human breast cancer cells into the superficial epigastric artery, which precludes the growth of metastases in body areas other than the respective hind leg1. Following 25-30 days after tumor cell inoculation, site-specific bone metastases develop, restricted to the distal femur, proximal tibia and proximal fibula1. Morphological and functional aspects of angiogenesis can be investigated longitudinally in bone metastases using magnetic resonance imaging (MRI), volumetric computed tomography (VCT) and ultrasound (US).
MRI displays morphologic information on the soft tissue part of bone metastases that is initially confined to the bone marrow cavity and subsequently exceeds cortical bone while progressing. Using dynamic contrast-enhanced MRI (DCE-MRI) functional data including regional blood volume, perfusion and vessel permeability can be obtained and quantified2-4. Bone destruction is captured in high resolution using morphological VCT imaging. Complementary to MRI findings, osteolytic lesions can be located adjacent to sites of intramedullary tumor growth. After contrast agent application, VCT angiography reveals the macrovessel architecture in bone metastases in high resolution, and DCE-VCT enables insight in the microcirculation of these lesions5,6. US is applicable to assess morphological and functional features from skeletal lesions due to local osteolysis of cortical bone. Using B-mode and Doppler techniques, structure and perfusion of the soft tissue metastases can be evaluated, respectively. DCE-US allows for real-time imaging of vascularization in bone metastases after injection of microbubbles7.
In conclusion, in a model of site-specific breast cancer bone metastases multi-modal imaging techniques including MRI, VCT and US offer complementary information on morphology and functional parameters of angiogenesis in these skeletal lesions.

In the pathogenesis of bone metastasis, angiogenesis is a crucial process and therefore represents a target for imaging and therapy. Here, we present a rat model of site-specific breast cancer bone metastasis and describe strategies to non-invasively image angiogenesis in vivo using magnetic resonance imaging, volumetric computed tomography and ultrasound.

Angiogenesis is an essential feature of cancer growth and metastasis formation. In bone metastasis, angiogenic factors are pivotal for tumor cell ...

Medicine

Models of Bone Metastasis

Bone metastases are a common occurrence in several malignancies, including breast, prostate, and lung. Once established in bone, tumors are responsible for significant morbidity and mortality1. Thus, there is a significant need to understand the molecular mechanisms controlling the establishment, growth and activity of tumors in bone. Several in vivo models have been established to study these events and each has specific benefits and limitations. The most commonly used model utilizes intracardiac inoculation of tumor cells directly into the arterial blood supply of athymic (nude) BalbC mice. This procedure can be applied to many different tumor types (including PC-3 prostate cancer, lung carcinoma, and mouse mammary fat pad tumors); however, in this manuscript we will focus on the breast cancer model, MDA-MB-231. In this model we utilize a highly bone-selective clone, originally derived in Dr. Mundy's group in San Antonio2, that has since been transfected for GFP expression and re-cloned by our group3. This clone is a bone metastatic variant with a high rate of osteotropism and very little metastasis to lung, liver, or adrenal glands. While intracardiac injections are most commonly used for studies of bone metastasis2, in certain instances intratibial4 or mammary fat pad injections are more appropriate. Intracardiac injections are typically performed when using human tumor cells with the goal of monitoring later stages of metastasis, specifically the ability of cancer cells to arrest in bone, survive, proliferate, and establish tumors that develop into cancer-induced bone disease. Intratibial injections are performed if focusing on the relationship of cancer cells and bone after a tumor has metastasized to bone, which correlates roughly to established metastatic bone disease. Neither of these models recapitulates early steps in the metastatic process prior to embolism and entry of tumor cells into the circulation. If monitoring primary tumor growth or metastasis from the primary site to bone, then mammary fat pad inoculations are usually preferred; however, very few tumor cell lines will consistently metastasize to bone from the primary site, with 4T1 bone-preferential clones, a mouse mammary carcinoma, being the exception 5,6.
This manuscript details inoculation procedures and highlights key steps in post inoculation analyses. Specifically, it includes cell culture, tumor cell inoculation procedures for intracardiac and intratibial inoculations, as well as brief information regarding weekly monitoring by x-ray, fluorescence and histomorphometric analyses.

Animal models are frequently utilized to study cancer metastasis to bone. In this protocol we will describe two common methods of tumor inoculation for bone metastasis studies and briefly describe some of the analyses utilized to monitor and quantify these models.

Bone metastases are a common occurrence in several malignancies, including breast, prostate, and lung. Once established in bone, tumors are responsible ...

Pathological Analysis of Lung Metastasis Following Lateral Tail-Vein Injection of Tumor Cells

Metastasis, the primary cause of morbidity and mortality for most cancer patients, can be challenging to model preclinically in mice. Few spontaneous metastasis models are available. Thus, the experimental metastasis model involving tail-vein injection of suitable cell lines is a mainstay of metastasis research. When cancer cells are injected into the lateral tail-vein, the lung is their preferred site of colonization. A potential limitation of this technique is the accurate quantification of the metastatic lung tumor burden. While some investigators count macrometastases of a pre-defined size and/or include micrometastases following sectioning of tissue, others determine the area of metastatic lesions relative to normal tissue area. Both of these quantification methods can be exceedingly difficult when the metastatic burden is high. Herein, we demonstrate an intravenous injection model of lung metastasis followed by an advanced method for quantifying metastatic tumor burden using image analysis software. This process allows for investigation of multiple end-point parameters, including average metastasis size, total number of metastases, and total metastasis area, to provide a comprehensive analysis. Furthermore, this method has been reviewed by a veterinary pathologist board-certified by the American College of Veterinary Pathologists (SEK) to ensure accuracy.

Intravenous injection of cancer cells is often used in metastasis research, but the metastatic tumor burden can be difficult to analyze. Herein, we demonstrate a tail-vein injection model of metastasis and include a novel approach to analyze the resulting metastatic lung tumor burden.

Metastasis, the primary cause of morbidity and mortality for most cancer patients, can be challenging to model preclinically in mice. Few spontaneous ...

Modeling Primary Bone Tumors and Bone Metastasis with Solid Tumor Graft Implantation into Bone

Primary bone tumors or bone metastasis from solid tumors result in painful osteolytic, osteoblastic, or mixed osteolytic/osteoblastic lesions. These lesions compromise bone structure, increase the risk of pathologic fracture, and leave patients with limited treatment options. Primary bone tumors metastasize to distant organs, with some types capable of spreading to other skeletal sites. However, recent evidence suggests that with many solid tumors, cancer cells that have spread to bone may be the primary source of cells that ultimately metastasize to other organ systems. Most syngeneic or xenograft mouse models of primary bone tumors involve intra-osseous (orthotopic) injection of tumor cell suspensions. Some animal models of skeletal metastasis from solid tumors also depend on direct bone injection, while others attempt to recapitulate additional steps of the bone metastatic cascade by injecting cells intravascularly or into the organ of the primary tumor. However, none of these models develop bone metastasis reliably or with an incidence of 100%. In addition, direct intra-osseous injection of tumor cells has been shown to be associated with potential tumor embolization of the lung. These embolic tumor cells engraft but do not recapitulate the metastatic cascade. We reported a mouse model of osteosarcoma in which fresh or cryopreserved tumor fragments (consisting of tumor cells plus stroma) are implanted directly into the proximal tibia using a minimally invasive surgical technique. These animals developed reproducible engraftment, growth, and, over time, osteolysis and lung metastasis. This technique has the versatility to be used to model solid tumor bone metastasis and can readily employ grafts consisting of one or multiple cell types, genetically-modified cells, patient-derived xenografts, and/or labeled cells that can be tracked by optical or advanced imaging. Here, we demonstrate this technique, modeling primary bone tumors and bone metastasis using solid tumor graft implantation into bone.

Bone metastasis models do not develop metastasis uniformly or with a 100% incidence. Direct intra-osseous tumor cell injection can result in embolization of the lung. We present our technique modeling primary bone tumors and bone metastasis using solid tumor graft implantation into bone, leading to reproducible engraftment and growth.

Primary bone tumors or bone metastasis from solid tumors result in painful osteolytic, osteoblastic, or mixed osteolytic/osteoblastic lesions. These ...

A Syngeneic Orthotopic Osteosarcoma Sprague Dawley Rat Model with Amputation to Control Metastasis Rate

The most recent advance in the treatment of osteosarcoma (OS) occurred in the 1980s when multi-agent chemotherapy was shown to improve overall survival compared to surgery alone. To address this problem, the aim of the study is to refine a lesser-known model of OS in rats with a comprehensive histologic, imaging, biologic, implantation, and amputation surgical approach that prolongs survival. We used an immunocompetent, outbred Sprague-Dawley (SD), syngeneic rat model with implanted UMR106 OS cell line (originating from a SD rat) with orthotopic tibial tumor implants into 3-week-old male and female rats to model pediatric OS. We found that rats develop reproducible primary and metastatic pulmonary tumors, and that limb amputations at 3 weeks post implantation significantly reduce the incidence of pulmonary metastasis and prevent unexpected deaths. Histologically, the primary and metastatic OSs in rats were very similar to human OS. Using immunohistochemistry methods, the study shows that rat OS are infiltrated with macrophages and T cells. A protein expression survey of OS cells reveals that these tumors express ErbB family kinases. Since these kinases are also highly expressed in most human OSs, this rat model could be used to test ErbB pathway inhibitors for therapy.

Here, a syngeneic orthotopic implantation followed by an amputation procedure of the osteosarcoma with spontaneous pulmonary metastasis that can be used for preclinical investigation of metastasis biology and development of novel therapeutics is described.

The most recent advance in the treatment of osteosarcoma (OS) occurred in the 1980s when multi-agent chemotherapy was shown to improve overall survival ...

An Orthotopic Murine Model of Human Prostate Cancer Metastasis

Our laboratory has developed a novel orthotopic implantation model of human prostate cancer (PCa). As PCa death is not due to the primary tumor, but rather the formation of distinct metastasis, the ability to effectively model this progression pre-clinically is of high value. In this model, cells are directly implanted into the ventral lobe of the prostate in Balb/c athymic mice, and allowed to progress for 4-6 weeks. At experiment termination, several distinct endpoints can be measured, such as size and molecular characterization of the primary tumor, the presence and quantification of circulating tumor cells in the blood and bone marrow, and formation of metastasis to the lung. In addition to a variety of endpoints, this model provides a picture of a cells ability to invade and escape the primary organ, enter and survive in the circulatory system, and implant and grow in a secondary site. This model has been used effectively to measure metastatic response to both changes in protein expression as well as to response to small molecule therapeutics, in a short turnaround time.

This orthotopic model of human prostate cancer allows for quantification of tumor size, circulating tumor cells, and formation of distinct metastasis to the lung. As cells must escape the primary organ, enter the blood stream, and implant into a secondary site, this model effectively recapitulates the scenario in humans.

Our laboratory has developed a novel orthotopic implantation model of  human prostate cancer (PCa). As PCa death is not due to the primary tumor, but  ...

Intratibial Osteosarcoma Cell Injection to Generate Orthotopic Osteosarcoma and Lung Metastasis Mouse Models

Osteosarcoma is the most common primary bone cancer in children and adolescents, with lungs as the most common metastatic site. The five-year survival rate of osteosarcoma patients with pulmonary metastasis is less than 30%.Therefore, the utilization of mouse models mimicking the osteosarcoma development in humans is of great significance for understanding the fundamental mechanism of osteosarcoma carcinogenesis and pulmonary metastasis to develop novel therapeutics. Here, detailed procedures are reported to generate the primary osteosarcoma and pulmonary metastasis mouse models via&#160;intratibia injection of osteosarcoma cells. Combined with the bioluminescence or X-ray live imaging system, these living mouse models are utilized to monitor and quantify osteosarcoma growth and metastasis. To establish this model, a basement membrane matrix containing osteosarcoma cells was loaded in a micro-volume syringe and injected into one tibia of each athymic mouse after being anesthetized. The mice were sacrificed when the primary osteosarcoma reached the size limitation in the IACUC-approved protocol. The legs bearing osteosarcoma and the lungs with metastasis lesions were separated. These models are characterized by a short incubation period, rapid growth, severe lesions, and sensitivity in monitoring the development of primary and pulmonary metastatic lesions. Therefore, these are ideal models for exploring the functions and mechanisms of specific factors in osteosarcoma carcinogenesis and pulmonary metastasis, the tumor microenvironment, and evaluating the therapeutic efficacy in vivo.

The present protocol describes intratibia osteosarcoma cell injection to generate mouse models bearing orthotopic osteosarcoma and pulmonary metastasis lesions.

Osteosarcoma is the most common primary bone cancer in children and adolescents, with lungs as the most common metastatic site. The five-year survival ...

Intra-Cardiac Injection of Human Prostate Cancer Cells to Create a Bone Metastasis Xenograft Mouse Model

As the most common male malignancy, prostate cancer (PC) ranks second in mortality, primarily due to a 65%-75% bone metastasis rate. Therefore, it is essential to understand the process and related mechanisms of prostate cancer bone metastasis for developing new therapeutics. For this, an animal model of bone metastasis is an essential tool. Here, we report detailed procedures to generate a bone metastasis mouse model via intra-cardiac injection of prostate cancer cells. A bioluminescence imaging system can determine whether prostate cancer cells have been accurately injected into the heart and monitor cancer cell metastasis since it has great advantages in monitoring metastatic lesion development. This model replicates the natural development of disseminated cancer cells to form micro-metastases in the bone and imitates the pathological process of prostate cancer bone metastasis. It provides an effective tool for further exploration of the molecular mechanisms and the in vivo therapeutic effects of this disease.

Here, we present a protocol for the intra-cardiac injection of human prostate cancer cells to generate a mouse model with bone metastasis lesions.

As the most common male malignancy, prostate cancer (PC) ranks second in mortality, primarily due to a 65%-75% bone metastasis rate. Therefore, it is ...

Machine Learning Algorithms for Early Detection of Bone Metastases in an Experimental Rat Model

Samenvatting

Meer video's verkennen