Interpretable deep learning model for distinguishing tumor pseudoprogression from true progression using MRI imaging of glioblastoma patients

Glioblastoma multiforme (GBM) are extremely invasive cancers. The treatment of GBM involves microsurgical resection followed by radiochemotherapy and chemotherapy. As a response to radiation treatment, in many cases, a new or a progressing lesion is observed in imaging studies which resolves without additional treatment. This phenomenon is called pseudoprogression (PsP). In contrast to PsP, a True Progression (TP) represents an enlarging lesion that requires a change in the treatment. Distinguishing between PsP and TP is thus central to treatment choice and clinical management. However, both types of progression present themselves with overlapping characteristics in imaging as assessed by radiologists. An automated machine learning method that can learn to discover distinctive markers to reliably differentiate between the two situations will thus be an effective prognostic tool in the clinical management of GBM. In this paper we present a 3D convolutional neural network (CNN) trained on 3D MRI images from 114 GBM patients to classify PsP and TP. Using a 5-fold cross validation strategy, we report multiple metrics by evaluating the model performance on left-out MRI image volumes not used during training. Specifically, our trained model performs with: AUCROC: 0.74, Peak geometric mean of specificity and sensitivity: 0.69, Brier Score: 0.22, Scaled Brier Score: 0.04. We also present decision curve analysis for our model. Prior works on this topic have reported only AUCROC. For model interpretability, we have used the technique GradCAM to discover and visualize the most salient regions in the MRI volume that are used by the CNN for making the decision. Our results show the neural network paying more attention to the lesion and peri-tumoral regions. These findings suggest further investigation of deep learning models trained on larger imaging datasets to build more robust and generalizable models for distinguishing between PsP and TP.