PATH-13. IMAGE-BASED PHENOTYPING OF GLIOMAS FOR CLINICAL USE. (14th November 2022)

• Record Type: Journal Article
• Authors: Becker, Aline; Casukhela, Ishya; Chan, Jany; McElroy, Joseph; Eicher, Tara; Fleming, Jessica; Haque, S Jaharul; Machiraju, Raghu; Chakravarti, Arnab
• Is Part Of: Neuro-oncology. Volume 24(2022)Supplement 7
• Page Start: vii152
• Page End: vii152
• Publication Date: 2022-11-14
• Subjects: Brain Neoplasms -- Periodicals; Brain -- Tumors -- Periodicals; Brain -- Cancer -- Periodicals; Nervous system -- Cancer -- Periodicals; 616.99481
• DOI: 10.1093/neuonc/noac209.586
• Languages: English
• ISSNs: 1522-8517

Abstract

Abstract: Adult-type diffuse gliomas have been classified according to histopathological characteristics only. Based on molecular profiles, the World Health Organization (WHO) classification defines three distinct biologic and prognostic types of diffuse gliomas: (i) IDH wild type ( IDH wt), (ii) IDH mutant, 1p19q intact ( IDH mut-non-codel), and (iii) IDH mutant, 1p19q codeleted ( IDH mut-codel). Gliomas with missing molecular information are classified as "Not otherwise specified" (NOS). A histopathological signature that distinguishes IDH status in gliomas using hematoxylin & eosin (H&E)-stained whole slide images (WSIs) has yet to be developed to be used as proxy for molecular alterations. We employed WSIs with corresponding molecular data from The Cancer Genome Atlas (TCGA) and reclassified the gliomas according to the 2021 WHO guidelines. Weakly supervised deep learning approaches, i.e. the Uncertainty-Aware CNN (UA-CNN), and accompanying workflows developed by our group were successful in the segmentation and classification of other cancers. We used the weak labels of IDHwt or IDHmut for given WSIs and utilized an informative sampling algorithm to identify the most relevant tiles that are predictive of their WSI-level label. Next, a second UA-CNN was trained on the relevant subset of tiles determined from the screening stage. For inferencing, given a WSI, the trained UA-CNN model yielded both a predicted label and uncertainty measure for each tile. When reassembled into a WSI, an aggregated uncertainty classification map is formed, which will allow neuropathologists to more rapidly identify regions of interest. While molecular testing is indispensable for current glioma classification, our tool may help reduce the number of NOS diagnosis, mainly in smaller services, by identifying image features that can differentiate between IDHwt and IDHmut. We expect to develop and deploy a method to increase the amount of information extracted from the H&E stain, and help prioritize downstream molecular analyses mandatory for the layered diagnosis of gliomas.