Uncertainty-assisted virtual immunohistochemical detection on morphological staining via semi-supervised learning

Abstract

Tumor suppressor gene TP53 plays a crucial role in cancer diagnosis and prognosis. The gene encodes the tumor suppressor protein p53, which can be identified through immunohistochemical (IHC) staining in various cancers, including gastric carcinoma. However, IHC staining is more costly and therefore not as prevalent as routine hematoxylin-eosin (H&E) staining. In this study, we present a semi-supervised learning-based approach for immunological detection (SSID) of TP53 mutation directly on H&E-stained gastric tissue sections, intending to improve gastric cancer diagnosis. SSID is trained on a small set of annotated image pairs and a larger unannotated dataset of H&E-stained images. It can detect the regions showing strong p53 expression, indicating TP53 mutation, and we validate the accuracy of our approach through both qualitative assessment (pathologists' average score of 2.22/3) and quantitative evaluation (e.g., averaged mean Intersection-over-Union of 0.73). Moreover, we introduce Bayesian uncertainty to assess the credibility of the detected masks, aiming to prevent misdiagnosis and inappropriate treatment. Our results demonstrate that SSID can circumvent the expensive and laborious IHC staining procedures and enable the diagnosis and prognosis of gastric cancer through immunological detection of TP53 mutation.