Multi-scale attention-based deep learning framework for tumor microenvironment profiling

The tumor microenvironment (TME) is an important factor in cancer development, treatment response, and immune regulation. Segmenting tumor subregions in histopathological images remains a challenge due to heterogeneity in space, morphology, and staining. In this regard, this paper presents a Histology-Guided Deep Learning Framework (HG-DLF) that involves multi-scale feature fusion, dual attention mechanisms, and graph convolutional networks to achieve accurate and robust TME analysis. Using the PanNuke dataset of 7,904 histology image patches across 19 tissue types, HG-DLF successfully segments tumor, immune, stromal, and nuclear structures. The model demonstrates a Dice Similarity Coefficient of 97.8%, an Intersection over Union (IoU) of 98.34%, and a Hausdorff Distance of 1.67—well surpassing baseline models, such as Deep CNNs and Her2Net, by more than 20% in accuracy and over 50% in inference time. The dual attention mechanism facilitates discriminative feature extraction, and the graph-based module leverages spatial context and tissue boundaries. The model demonstrates greater generalizability across folds in k-fold cross-validation and high accuracy despite morphological differences. HG-DLF offers an interpretable, scalable computational pathology solution with applications for estimating tumor heterogeneity, immune infiltration levels, and supporting clinical decision-making in precision oncology.