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

Summary

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.

Abstract

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’s tumor-take rate is 60%–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.

Introduction

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 disease^1,2. This initial invasion can be considered an early step in metastatic disease, but is typically missed during conventional staging examinations^3,4. 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 used^6,7,8. 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%–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.

Protocol

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 elsewhere^6,8. The most relevant steps are presented below.

1. 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 °C, 5% CO₂) and passage the cells 2–3 times a week.
2. 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’s chamber and resuspend them in 200 µL of RPMI-1640 at a concentration of 1.5 x 10⁵ cells/200 µL.
3. Use 6–8 week-old nude rats and keep them under pathogen-free, controlled conditions (21 °C ± 2 °C room temperature, 60% humidity, and 12 h light-dark rhythm). Offer autoclaved feed and water ad libitum.
4. Before performing the surgery, inject an analgesic drug (e.g., Carprofen 4 mg/kg) subcutaneously. Anesthetize rats with an isoflurane (1–1.5 vol. %)/oxygen mixture at a flow rate of 2 L/min. Check the anesthetic depth by toe pinching.
5. For surgery, use an operating microscope with a 16x magnification.
6. Perform a 2–3 cm cut in the rat’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.
7. Ligate the distal portion of the SEA. Perform a cut of the SEA’s wall and insert a 0.3 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.
8. Slowly inject the MDA-MB-231 cell suspension from step 1.2 (1.5 x 10⁵ cells/200 µL) 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.

2. Magnetic resonance imaging (MRI)

NOTE: For a detailed description of MRI procedures, please see Bäuerle et al.¹¹.

1. Perform MRI 10 days PI using a dedicated experimental scanner (see Table of Materials) or a human MR system with an appropriate animal coil.
2. Anesthetize the rat with an isoflurane (1–1.5 vol. %)/oxygen mixture as described above. Place a catheter in the rat’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).
3. 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 mm; scan time 11:24 min).
4. 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 mm²; slice thickness = 1 mm; 8 slices; 100 time points; scan time = 8:25 min). After 30 s, start injecting the contrast agent over a time period of 10 s.
   NOTE: The total time to perform an MRI examination is approximately 20 min per animal.

3. Positron emission tomography/computed tomography (PET/CT)

NOTE: For a detailed description of the PET procedures, please see Cheng at al.¹².

1. Perform PET/CT imaging 10 days PI using a dedicated experimental scanner (see Table of Materials).
2. 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.
3. Inject 6 MBq of ¹⁸F-Fluorodeoxyglucose (¹⁸F-FDG) into the tail vein and wait ~30 min to allow the tracer to distribute properly.
4. Perform a CT acquisition (tube voltage = 80 kV, tube current = 500 µA, isotropic resolution = 48.9 µm, duration = 10 min).
5. Perform a static PET acquisition (lower/upper discriminatory level = 350/650 keV; timing window = 3.438 ns; duration = 15 min).

4. Alternative imaging strategies

1. For an early assessment of MDA-MB-231 cells in the hind leg, inoculate 1.5 x 10⁵ labeled cells /200 µL 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. 
2. Perform experimental ultrasound using a dedicated scanner after intravenous injection of microbubbles to derive morphological and functional parameters of vascularization comparable to MRI⁷.

5. MRI analysis

1. Use a DICOM viewer¹⁵ with a DCE Plugin¹⁶ and load the DCE sequence in 4D-mode by clicking the “Import” button in the top menu, selecting the DICOM folder containing the MR images from step 2.4, and clicking “4D Viewer” in the top menu.
2. Place a circular 2-dimensional region of interest (ROI), with a target size of 1.5 mm², in the proximal tibial shaft’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.
3. Start the DCE plugin from the top menu, select “Relative Enhancement” in the “Plot Type” 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 “DCEraw.txt” as the file name.
4. Open RStudio¹⁷ and load the provided DCE-Script.R file via the “File” menu by selecting “Open File”. Run the entire script by selecting “Code”, then “Run Region” and then “Run All” from the menu. Copy the output to the provided template file named “ImagingFeatures.xlsx” (Figure 2).
5. In the DICOM viewer, place a second ROI within the back muscle of the animal and repeat steps 5.2–5.4 to obtain the muscle DCE measurements for normalization purposes. Within the spreadsheet “ImagingFeatures.xlsx”, the respective bone measurements are automatically divided by the respective muscle measurements for normalization purposes.
6. Repeat steps 5.1–5.5 for all animals and complete the spreadsheet.

6. PET/CT analysis

1. Open the PET/CT analysis software and import the data obtained in step 3 by clicking “File”, followed by “Manual import”. Mark the ct.img.hd and the pet.img.hdr files. Click “Open” and select “Import all”.
2. Open the datasets by selecting “General analysis”, followed by “OK”.
3. Select “ROI Quantification”, followed by “Create”, and then “Create a ROI from a template”. Place a 2-dimensional ROI approximately 4 mm x 6 mm into the proximal tibial shaft’s bone marrow of the right hind leg.
4. Select “ROIs (Target 1 overlay)” and write down mean, minimum, and maximum values in Bq/mL.
5. Calculate the maximum standardized uptake value (SUV_max): 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).

7. Determining the tumor-take rate

1. To diagnose tumor growth in the right hind leg, repeat MR and PET/CT imaging on day 30 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 SUV_max in PET/CT. According to previous experiments, 60%–80% of the animals will develop metastases in their right hind leg.
2. Complete the spreadsheet by adding an additional “Tumor” column and enter “1” for every animal that presents metastases, and “0” for every animal without visible tumor burden (Figure 2). Save the spreadsheet as “ImagingFeatures.xlsx” within the Downloads folder.

8. Feature selection

1. 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.
2. Draw the File-subroutine from the Data menu into the workspace on the right and double-click it. Load the spreadsheet by clicking the “Folder” icon and selecting the file “ImagingFeatures.xlsx”. Select the “Export” worksheet and assign the target-attribute to the variable “Tumor”. Assign the “Skip” function to the animal number (Figure 3).
3. Draw the “Rank” subroutine from the Data menu into the workspace and connect the “File” and “Rank” subroutines by drawing a line between them.
4. Open the “Rank” subroutine by double-clicking on its icon, and select the “Information Gain” algorithm¹⁹.
5. From the five acquired parameters, use the top three for further analyses (SUV_max, PE, and AUC).
   Note: These parameters reflect metabolic activity (SUV_max) and tissue vascularization (PE and AUC).

9. ML analysis

1. Open RStudio 3.4.1¹⁷ and load the provided TrainModel.R-Script via the “File” menu.
2. Install the required libraries (this only has to be done once) by typing: install.packages(c("caret", "readxl", "pROC", "RcmdrPlugin.EZR", "ggplot2"))
3. To load the required libraries and set the Downloads folder as the working directory, select the lines 3–5 within the TrainModel.R Script.
4. Run the selected code by clicking “Code” within the menu, and then “Run Selected Line(s)”.

10. Training an avNNet ML algorithm

1. To train an avNNet algorithm, select the lines 8–39 from the TrainModel.R-Script (see step 9.1).
2. Run the selected code by clicking “Code” within the menu, and then “Run Selected Line(s)”.

11. Analyzing the ML algorithm’s results

1. To assess standard parameters of diagnostic accuracy (sensitivity, specificity, positive and negative predictive values, and likelihood ratios), select the lines 41–50 from the TrainModel.R-Script.
2. Run the selected code by clicking “Code” within the menu, and then “Run Selected Line(s)”.

12. Comparing the final model's Receiver Operating Characteristic (ROC) curve with the ROC curves of its constituent parameters

1. To perform DeLong’s tests to compare the model’s ROC curve with the ROC curves of its constituent parameters, select the lines 52–62 from the TrainModel.R-Script (see step 9.1).
2. Run the selected code by clicking “Code” within the menu, and then “Run Selected Line(s)”.

Results

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’s right proximal tibia is presented in Figure 2A. The DCE raw measurements were saved by selecting the “Export” button and choosing “DCEraw.txt” as the file name.

Subsequent calculations of the dynamic p...

Discussion

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 c...

Materials

List of materials used in this article

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