Pearce, Thomas M.

Abstract Computer vision now enables deep learning models to link histopathologic features to molecular features in a quantitative manner not previously achievable by human pathologists. However, human interpretability is often limited. We trained a deep neural network to predict gene expression from digital whole slide images (WSIs) using 848 central nervous system (CNS) tumors with paired WSI and RNA-sequencing data. We used inferred RNA expression levels as a surrogate for protein expression of the CNS tumor lineage markers GFAP, OLIG2, and SSTR2, as well as the proliferation marker MKI67 (Ki67). We established thresholds for categorizing the inferred expression levels as positive or negative based on levels observed in cross-validation testing in select tumor types to approximate routine clinical interpretation of immunohistochemical (IHC) staining. We tested the sensitivity and specificity of our ‘virtual’ IHC on an independent, multi-institutional cohort of over 2,000 CNS tumors with objective diagnostic labels derived from DNA methylation-based tumor classification. As a dichotomous variable, inferred GFAP expression showed a sensitivity of 74% and a specificity of 99% as a glial marker in a cohort of gliomas and meningiomas. In the same cohort, OLIG2 showed a sensitivity of 75% and a specificity of 97% as a glial marker, while SSTR2 showed a sensitivity of 83% and a specificity of 97% as a marker of meningioma. In a cohort of ependymomas and gliomas, virtual IHC for OLIG2 showed a sensitivity of 75% and a specificity of 82% in distinguishing ependymomas from gliomas. Finally, in a cohort of gliomas, inferred expression of MKI67 demonstrated a trend consistent with what would be expected by actual IHC, with increasing MKI67 expression with increasing tumor grade. Our model provides the basis for a human interpretable and clinically applicable deep neural network to aid human pathologists in the diagnosis and grading of CNS tumors.

Abstract Molecular testing has become critical components in the diagnostic classification of the central nervous system (CNS) tumors. However, these methods require substantial resources, limiting accessibility for many patients. Recent advances in artificial intelligence (AI) and computer vision empower deep learning models to infer molecular features from histopathology images, which may often be more readily available than molecular testing, to classify CNS tumors. We trained pan-CNS tumor models to predict DNA methylation and gene expression from whole-slide images (WSIs). These molecular predictions were then used in a hierarchical machine learning framework, termed Neuropath-AI, to predict nine broad tumor families and 52 tumor types using a large diverse cohort of 5,715 histopathology samples, of which 2,988 had paired DNA methylation profiling and 848 had paired RNA-sequencing. We evaluated Neuropath-AI on another large independent multi-institutional cohort of 5,516 CNS tumors, for which Neuropath-AI predicted tumor class with an associated confidence score. Neuropath-AI designated 46% of test samples as predictable with high-confidence, which it then accuracy classified in 97% of samples. It made moderate-confidence (and above) predictions for 87% of test samples, for which it achieved a diagnostic accuracy of 80% for the top-1 prediction and 86% accuracy for a top-2 predictions. A comparison with human neuropathologists showed that Neuropath-AI is comparable at classifying tumors. A second test showed that human neuropathologists were more accurate when the Neuropath-AI classifier results were provided to them, testifying to the potential translational benefit of integrating AI classification tools into the pathologist’s workflow. Our model provides the basis for a clinically applicable deep learning assistant to improve human efficiency and accuracy of CNS tumor diagnosis. The model will be made publicly available and can be readily implemented to assist human pathologists in clinical workflows, in further expanded prospective studies.