Automatic Lymphocyte Detection on Gastric Cancer IHC Images Using Deep Learning

Gastric cancer remains one of the deadliest malignancies globally, with poor prognosis often attributed to late-stage detection and inconsistent pathological interpretation. This paper proposes an Integrated Gastric Cancer AI Diagnostic System (IGCAIDS), an end-to-end deep learning pipeline for automated analysis of scanned histopathological reports, specifically Hematoxylin and Eosin (H&E) stained whole-slide images (WSIs) and Immunohistochemistry (IHC) images, to accurately classify whether a patient has gastric cancer and predict clinical outcomes. Building upon three landmark studies, namely Wang et al. (2021) on lymph node WSI analysis via ResNet-50, Garcia et al. (2017) on tumor-infiltrating lymphocyte (TIL) detection from IHC images using a Deep Convolutional Neural Network (DCNN), and Choi and Kim (2023) on the AI pathology landscape in gastric cancer, we synthesize a novel six-stage pipeline that: (1) ingests a scanned pathological report image, (2) performs stain normalization and preprocessing, (3) segments tissue regions using U-Net, (4) classifies cancerous versus non-cancerous areas using ResNet-50, (5) detects and quantifies TILs, and (6) computes the tumor-area-to-metastatic-lymph-node-area ratio (T/MLN) and generates a structured diagnostic report with a binary cancer verdict, confidence score, and prognostic stratification. The framework achieves 97.2% accuracy, 98.5% sensitivity, 96.5% specificity, and AUC of 0.992 in AI-only mode. The T/MLN ratio, proven to be an independent prognostic indicator with a hazard ratio of 2.05 (95% CI: 1.66-2.54, p