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

This protocol facilitates the development of a machine learning algorithm for predicting the growth of bone metastases in an experimental model at the stage of early organ colonization. The main advantage of this technique is that it combines several imaging parameters into a machine learning algorithm that significantly outperforms the productive ability of each individual parameter. Though, this protocol aims for the early diagnosis of metastases in bone, it can be adapted to different organs or areas of multimodal and multiparametric imaging research.

10 days post-surgery and injection in a DICOM viewer with a DCE plugin. Pick import to load the DCE sequence in 4D mode and select the DICOM folder containing the MRI images. Place a circular 1.5 square millimeter to dimensional region of interest in the proximal tibial shaft bone marrow of the right hind leg within the fourth or fifth MRI image and select relative enhancement in the plot type.

Enter the baseline range from time points, one to five into the respective fields and export the analysis as a dot TXT file named DCE raw dot TXT. Open RStudio and load the provided DCE script dot R file. Select code, run region and run all to run the entire script and copy the output to the provided imaging features template file.

In the DICOM viewer, place a second region of interest within the back muscle of the animal and repeat the DCE measurements as demonstrated. In the imaging feature spreadsheet the respective bone measurements will be automatically divided by the respective muscle measurements for data normalization. To analyze the PET/CT images open the PET/CT analysis software and import the PET/CT images.

Click general analysis and select region of interest quantification, create and create a region of interest from a template. Place a four by six millimeter, to dimensional region of interest within the bone marrow of the proximal tibial shaft of the right hind leg and select regions of interest target one overlay. Note the mean, minimum and maximum values in becquerels per milliliter.

Then, divide the maximum value by the injected activity and multiply the result by the animal weight in grams to calculate the standardized uptake value. To diagnose the tumor growth rate in the injected hind leg. After obtaining MRI and PET/CT images on day 30 post-injection, analyze the images as demonstrated and add a tumor column to the imaging features spreadsheet.

Then enter a one or zero for animals with metastasis or without a visible tumor burden respectively. To determine the most relevant features for the prediction of future tumor growth, import the spreadsheet into an open source data visualization machine learning and data mining toolkit. Drag the file sub-routine from the data menu into the workspace and double-click the file.

Click the folder icon to load the spreadsheet and select the imaging features spreadsheet. Select the export worksheet and assign the target attribute to the variable tumor. Assign the skip function to the animal number.

Drag the rank sub-routine into the workspace and draw a line to connect to the file and rank sub-routines. Then, double click to open the rank sub-routine and select the information gain algorithm. For machine learning analysis, open RStudio 3.4.1 and load the provided train model R script.

Select lines three through six within the script to load the required libraries. Click code to run the selected code and click run selected lines to run the selected lines. To train a model average to neural network algorithm, select lines eight through 39 from the train model R script and click code and run selected lines.

Then to assess the standard parameters of diagnostic accuracy, select lines 41 through 50 from the train model R script and click code and run selected lines. On day 10 after surgery and cancer cell injection, MRI and PET/CT images can be acquired. DCE analysis allows the measurement of muscle and bone tissue areas of interest.

These values can be normalized by dividing the bone measurements by the muscle measurements. On day 30 post-injection, all of animals are evaluated to determine whether they have developed metasticies with a one indicating a positive tumor burden and a zero indicating healthy animals without visible tumors. Running the train model R script allows the optimal hyper parameter combination to be determined and the final model to be calculated using the optimal hyper parameter combination.

These data allow a set of standard diagnostic parameters to be calculated and the receiver operating characteristic curve of the model to be plotted. For example, in this analysis of 28 samples the model performs significantly better than all of its three constituents. Several machine learning algorithms tend to perform better when the input data are normalized.

In this protocol normalization is achieved using the Box Cox function. This protocol uses a model average during network as a machine learning algorithm. However, the provided framework can easily be adapted to other algorithms such as random forests or support vector machines.

Extracting numerical information from image material has become essential. Algorithms such as these may facilitate the integration of large amounts of data to allow patient stratification.