Purpose: Glioblastoma (GBM) is the most common primary brain tumor with an overall survival (OS) of only 16 months. Early survival prediction is essential for treatment decision-making. Here we proposed a reference-guided contrastive learning approach to predict GBM patient post-operative OS using only pre-operative MRIs.
Methods: Our dataset was from the Brain Tumor Segmentation (BraTS) challenge 2020 consisting of multimodal pre-operative MRI scans of 235 GBM patients with survival days. The challenge task was to classify GBM patients into 3 classes: long-survivors, mid-survivors and short-survivors. The backbone of our approach was a Siamese network consisting of twinned ResNet-based feature extractors followed by a 3-layer classifier. During training, the feature extractors explored the traits of intra and inter-class by minimizing the contrastive loss of randomly paired 2D MRIs, and the classifier utilized the extracted features to generate labels with cost defined by cross-entropy loss. During testing, the extracted features were also utilized to define distance between the test sample and the reference composed of training data, to generate an additional predictor via K-NN classification. And the final label was the ensemble classification from the network and K-NN.
Results: Following the BraTS challenge protocol, we assessed the accuracy (ACC) of 3-class classification, the mean (MSE), median and standard deviation of the square error of survival days, which are 65%, 74220, 9418 and 120403, respectively.
Conclusion: Predicting GBM patient OS using only pre-operative MRIs is challenging. Our reference-guided contrastive learning approach demonstrated promising ability in mining discriminative features, and achieved a higher ACC of 65% and much smaller MSE of 74220 compared to the top-ranked result from the BraTS challenge 2020 (Rank 1st: 61%, 391589) with minimal manual processing and generalization requirement. This prediction strategy can be potentially applied to assist clinical decision-making with objective and reproducible measures.
Funding Support, Disclosures, and Conflict of Interest: This work is supported by NIH R01 CA235723
TH- Response Assessment: Modeling: other than machine learning