Multi-Task Learning Model for Enhancing Tumor-Infiltrating Lymphocyte Prediction

Quantifying tumor-infiltrating lymphocytes (TILs) from histopathology images provides essential insights into tumor–immune interactions and serves as a key biomarker for prognosis and immunotherapy response in breast cancer. Conventional computational approaches require separate tissue segmentation and lymphocyte detection, which leads to redundant computation and limited information sharing between related features. To address these challenges, we propose a unified multi-task learning (MTL) framework that simultaneously performs tissue segmentation and lymphocyte detection to enhance TIL prediction accuracy and efficiency. The proposed model is built upon a pathology-specific large vision model (LVM) pretrained through self-supervised learning and incorporates layer-wise embedding sharing to enhance deep cross-task feature interactions. Experimental results on the TIGER Challenge dataset demonstrated that the proposed unified MTL model achieved improved TIL prediction accuracy compared to both single-task and conventional MTL models. In particular, compared to the conventional MTL model, the proposed approach achieved a 7.0% improvement in Pearson correlation and a 5.9% increase in Spearman correlation, while reducing the number of trainable parameters by approximately 33%, demonstrating superior performance in both predictive accuracy and computational efficiency. When benchmarked against top-performing models from the TIGER Challenge, the proposed framework achieved comparable or higher accuracy across tissue segmentation, lymphocyte detection, and TIL prediction metrics. The framework also improved concordance between predicted and ground-truth TIL scores across multiple correlation metrics while maintaining robust segmentation and detection performance. These results highlight that joint optimization of tissue- and cell-level tasks can yield biologically coherent representations of the tumor microenvironment, supporting automated TIL quantification. The proposed MTL approach provides a scalable framework with potential applicability in digital pathology and computational oncology.