Multi-scale Whole Slide Image Assessment Improves Deep Learning based WHO 2021 Glioma Classification
Keywords: glioma, multi-scale, deep learning, computational pathology
Abstract: The 2021 WHO classification of tumors of the central nervous system necessitates the integration of molecular and histologic profiling for a conclusive diagnosis of glioma. Molecular profiling is time-consuming and may not always be available. We hypothesize that sub-visual cues in whole slide images (WSI), not perceivable by the naked eye, carry a predictive value of molecular characteristics and can allow categorization of the adult infiltrative gliomas in one of three major types: i) oligodendroglioma, ii) astrocytoma, and iii) glioblastoma. Towards this end, we present a computational pipeline comprising patch analysis of Hematoxylin and Eosin (H\&E)-stained WSIs, feature encoding with ImageNet pretrained ResNet50, and an attention-based multiple instance learning paradigm. We trained individual models at four distinct magnification levels (20x, 10x, 5x, 2.5x), and assessed the fusion of various ensemble combinations to mimic the WSI assessment by expert pathologists, to capture local and global context. Our results using a multi-scale approach demonstrate 3-9\% improvement in classification accuracy when compared with models utilising a single magnification level. This advancement underscores the efficacy of attention-based models combined with multi-scale approaches in augmenting traditional assessment of WSIs. The implications of our findings are significant in enhancing glioma diagnosis and treatment planning in neuro-oncology, by enabling diagnostics in low-resource environments where molecular profiling is not available.
Submission Number: 14
Loading