Attention-driven multi-scale analysis for accurate tumor classification in digital pathology.

e13703 Background: The analysis of Whole Slide Images (WSIs) in digital pathology faces significant challenges due to their hierarchical structure and gigapixel-scale resolution. Traditional Convolutional Neural Networks struggle to simultaneously capture both tissue-level patterns and cellular-level features, potentially compromising diagnostic accuracy. This limitation is particularly critical in complex assessments like lung adenocarcinoma subtype classification, where accurate diagnosis requires integrating information across multiple spatial scales. While attention-based architectures show promise, their application to WSI analysis remains limited by computational constraints. Additionally, pathologists' increasing workloads necessitate more sophisticated computational methods that can support rapid and accurate diagnosis while maintaining interpretability. Methods: We propose a multi-scale analysis framework integrating the Swin Transformer with combining hierarchical shifted window partitioning using token learning. Our approach enables simultaneous analysis of cellular and tissue-level patterns. This design enhances both diagnostic accuracy and computational efficiency for clinical implementation. We validated our approach across three key histopathological datasets: CAMELYON16, WSSS4LUAD, and BMRIDS. For CAMELYON16, we employed 250,296 patches from 170 WSIs for normal/tumor classification training, with testing on 130 WSIs. The WSSS4LUAD dataset enabled subtype classification (Pure Tumor, Pure Stroma, Pure Normal, Tumor + Stroma, Tumor + Normal, All) using 10,093 patches. Results: Our most extensive experiment utilized the BMRIDS lung cancer dataset, involving five adenocarcinoma subtype classifications (lepidic, acinar, papillary, micropapillary and solid) across 1,074,785 patches from 143 WSIs. The model demonstrated exceptional performance, achieving over 90% accuracy in both binary and multi-class patch-level classifications within 50 epochs without parameter tuning. The model’s robustness across cancer types underscores its clinical potential for rapid histopathological assessment and precise subtype classification. Conclusions: Our work demonstrates that combining the Swin Transformer's hierarchical structure with adaptive token learning effectively addresses WSI analysis challenges in digital pathology. The framework's consistent high performance across diverse datasets validates its clinical potential, offering a possible path for the creation of effective and comprehensible computational tools for cancer classification and diagnosis according to WHO.