Work

A Hybrid CNN-Transformer Model for Tumor-Infiltrating Lymphocyte Score Prediction in Breast Cancer Histopathological Image

Year: 2026

Type: article

Abstract: Tumor-Infiltrating Lymphocytes (TILs) are crucial biomarkers in breast cancer, reflecting the immune response and providing prognostic information. Accurate quantification of TILs from whole-slide histopathology images remains challenging due to large image sizes, heterogeneous cellular distribution, and subtle morphological variations. To address these limitations, this study introduces a hybrid deep learning model, combining ResNet-152 and Vision Transformer (ViT) architectures, for automated TIL score prediction using the TIGER WSITILS dataset. Unlike conventional CNN-based approaches, which primarily capture local texture features, or transformer-only methods, which often lose fine-grained spatial detail, the proposed hybrid model leverages the strengths of both architectures—ResNet152 for rich local morphological representation and ViT for modeling long-range spatial relationships across tissue context. The model achieved a Mean Absolute Error (MAE) of 4.64, Mean Squared Error (MSE) of 48.21, and Root Mean Squared Error (RMSE) of 6.94 for regression-based prediction. In classification, it reached 94.74% accuracy with Pearson and Spearman correlation coefficients of 0.9894 and 0.9671, respectively. These results demonstrate that the proposed ResNet152-ViT hybrid framework effectively bridges local and global feature learning, offering improved accuracy, robustness, and interpretability for TIL assessment in breast cancer histopathology.

Source: Engineering Technology & Applied Science Research

DOI: 10.48084/etasr.15757

Authors: G. K. Shruthi, Pushpa Ravikumar

Institution: Visvesvaraya Technological University

Language: English

Topic: AI in cancer detection

Subfield: Artificial Intelligence

Field: Computer Science

Domain: Physical Sciences

SDG: Good health and well-being

Open Access status: gold

Estimated APC paid: $431