Pathologista-interpretable prediction of low grade glioma and glioblastoma from whole slide images

Classification of glioma involves both molecular and histological factors. Deep models utilizing whole slide images (WSIs) for a wide variety of classification tasks in glioma achieve high performance, but suffer from a lack of interpretability, hindering their adoption into clinical practice. Here, we apply a novel pretraining method that integrates pathologist interpretable concepts, increasing model interpretability while maintaining performance. Using this model in the supervised setting, we achieve an average AUC of 97.33 across 5 folds using an interpretable model, compared to an average AUC of 98.12 using an uninterpretable deep model. Using the concepts alone, we obtain an average AUC of 92.65. Use of an unsupervised pathologist heuristic-guided method for prediction achieves an average AUC of 80.08. We also provide an additional module to visualize concept expression in the most important WSI patches that influenced the model's decision. These findings highlight that concept-guided pretraining can bridge the gap between interpretability and performance, advancing the clinical translation of deep learning for glioma classification.