Deep learning-based approach to predict multiple genetic mutations in colorectal and lung cancer tissues using hematoxylin and eosin-stained whole-slide images.

• DOI: 10.1200/jco.2023.41.16_suppl.1549
• Publication Date: 2023-06-04T14:58:30Z
• Authors: Teppei Konishi; Mateusz Grynkiewicz; Keita Saito; Takuma Kobayashi; Akiteru Goto; Michinobu Umakoshi; Takashi Iwata; Hiroshi Nishio; Yuki Katoh; Tomonobu Fujita; Tomoya Matsui; Masaki Sugawara; Hiroyuki Sano
• Keywords: Microsatellite Instability; Digital Pathology; Stain

1549 Background: The presence of genetic mutations is a vital prognostic in many types cancer. However, genomic testing expensive and challenging to perform. In contrast, hematoxylin eosin (H&E) staining relatively inexpensive straightforward. Thus, this study, we propose method predicting the using H&E-stained whole-slide images (WSIs). Methods: We divided each H&E–stained WSI into small pieces or “patches.” used deep learning model classify patch based on tumor-containing regions. then extracted image features from learning-based feature extractor. created for entire by concatenating patches. trained mutation classification models as input absence output. Finally, evaluated performance these area under receiver operating characteristic curve (AUC). Results: First, our methods Cancer Genome Atlas (TCGA) colorectal cancer dataset. WSIs data associated with Microsatellite Instability ( MSI) BRAF gene mutations, which are directly relevant therapeutic strategies, obtained an independent clinical cohort 566 patients TCGA colon rectum adenocarcinoma. training, validation, test splits, comprising 367, 90, 109 patients, respectively. training validation splits selection, split evaluation. AUC values 95% confidence intervals (CIs) were 0.721 (CI = 0.572–0.870) MSI 0.712 0.547–0.877) mutations. also applied approach MUC16, KRAS, ALK lung 909 adenocarcinoma squamous cell carcinoma 582, 146, 181 contrast those dataset, generated all 0.897 0.85–0.95) 0.845 0.75–0.94) 0.756 0.57–0.94) Conclusions: proposed predict only its datasets. Our has potential certain superior performance. These predictions can be improve accuracy prediction alone.