M3CRNN - Predicting MGMT methylation from MRI scans

MRI to MGMT: Predicting Methylation Status in Glioblastoma Patients Using Convolutional Recurrent Neural Networks

Lichy Han and Maulik R. Kamdar
Program in Biomedical Informatics, Stanford University

This is the visualization platform for the Computational Recurrent Neural Network pipeline deployed to predict MGMT methylation status from Brain MRI scans of Glioblastoma Multiforme patients. This platform is deployed using Python Flask and Tensorflow, and can be used to click and load each MRI scan and visualize filter outputs. This platform is configurable for any volumetric biomedical data set, as well as a different CRNN pipeline. We also visualize the prediction results, as well as MRI scans.

Visualize the CRNN Filter Outputs

This view allows the domain user to load each MRI scan from the Training, Validation and Testing sets individually into our pre-trained CRNN pipeline. Once the pipeline computes the prediction score, each control (b) can be clicked to visualize the output of each filter and its ReLU activated output, in each convolutional layer. We believe that such a visualization is a first step to make deep learning models clinically intepretable. Each visualization opens in its own separate dialog that can be dragged around or closed, as desired. Moreover, the visualization is easily configurable for increasing the number of convolutional layers or different numbers of filters.

Visualize Prediction results

This view allows the domain user to see the classifier predictions over the test set. Each MRI scan in the test set has been divided into four groups --- True Positives, False Positives, False Negatives and True Negatives. The user can click and select each MRI scan and visualize the scan. Generally, features such as ring enhancement lesions, textures, location of the tumor, type of the MRI scan, shape of the tumor, stand out across the four groups, and generalized patterns can be generated upon comparison.

Visualize MRI Scans

This view allows the domain user to visualize the entire MRI scan, for each patient in the Training, Validation and Testing datasets. It was also previously used for annotating the different MRI scans based on the location of the tumor, as well as the start and end of the actual MRI scan (since there might be noisy pixels at the end or recurrent image frames). The domain user also can establish whether a given scan is valid or not (Since some MRIs are entirely black or white). The screenshot can be seen for further insight. We have disabled the annotation facility currently.