3593 - Assessing the Prognostic Value of Image Embeddings Extracted from a Convolutional Neural Network Trained for Prediction of IDH Status from H&E Images

Purpose/Objective(s): Glioblastoma (GBM) is one of the most complex and lethal brain cancers, accounting for 51% of primary malignant tumors. Gliomas were originally graded using histopathological features and were further refined in the 2021 WHO classification update to include molecule profiling. However, a visual signature for these molecular alterations has yet to be established. Machine vision models compute numerical representations of image data (i.e., embeddings), which can provide novel insight into disease states, particularly in digital pathology. We hypothesize that these deep learning (DL) methods can learn novel image features from H&E-stained WSIs of surgically resected tumors. Here, our aim was to improve our baseline model in order to (1) characterize molecular alterations of gliomas, (2) identify regions of interest in WSIs to assist pathologists in forming a diagnosis, and (3) improve patient stratification and prediction of patient outcomes.

Materials/Methods: The Cancer Genome Atlas (TCGA) database was filtered for patients (N = 463) with formalin-fixed paraffin-embedded tissue slides and molecular data. Sample were reclassified according to the 2021 WHO system as needed. Tumor tiles were extracted from expert-annotated WSIs at 20x magnification, resulting in a dataset of 3.7 million tiles without resampling or augmentation. Tile-level image embeddings were generated from a pre-trained self-supervised contrastive learning neural network, and multiple IDH predictors were trained using a planning tool dropout. The top performing models were used to build ensemble models for WSI-level IDH prediction. Post-hoc analyses were run on the extracted image embeddings to assess patterns in similarity, tissue segmentation, and prognostic value on overall survival.

Results: Using 5-fold, leave-one-out cross validation, we show notable improvements in performance metrics from our baseline model (precision = 0.84, sensitivity = 0.79, specificity = 0.78, negative predictive value = 0.81). Aggregating models into an ensemble and/or uncertainty weighting modestly increased tile-level performance but resulted in a significant improvement at the WSI-level (N = 256 WSIs, precision = 0.94, sensitivity = 0.97, specificity = 0.97, negative predictive value = 0.95). Clustering the image embeddings by cosine similarity revealed CNN-dependent regions of interest, which were evaluated for relevancy in predicting overall survival.

Conclusion: Ensemble DL models were successfully trained for IDH classification to a high degree of sensitivity and specificity at the level of WSIs. Analyses of the learned image embeddings suggest the presence of two regions of interest to facilitate tissue segmentation. Efforts are ongoing to expand the training dataset and to validate these results.