Enhancing Lung Cancer Histopathological Subtyping via Fuzzy Patch Scoring Integrated into an Ensemble Deep Learning Framework

Mohammad mehdi hosseini, Bardia Yousefi Rodd

Confirmation: I have read and agree with the IEEE BHI 2025 conference submission's policy on behalf of myself and my co-authors.

Keywords: Digital pathology, Latent space embedding, Ensemble model, Fuzzy scoring, Transfer learning

Abstract:

Automated histopathological subtyping of lung cancer using stained whole-slide images (WSIs) remains a pivotal yet formidable challenge, owing to pronounced tumor heterogeneity, intricate cellular morphology, and severe class imbalance within available datasets. Deep learning architectures vary in their capacity to capture diverse pathomic features, and the diagnostic efficacy of such models is heavily influenced by the quality of the tissue patches extracted from WSIs.

To address these limitations, we present a novel ensemble deep learning framework enhanced by a fuzzy-weighted patch quality assessment, which improves the selection and contribution of informative regions within WSIs. High-quality patches are identified using a fuzzy scoring method and processed through multiple pre-trained CNN and vision transformer (ViT) models to capture diverse feature representations. These features are integrated via latent embeddings, with fuzzy scores incorporated both as auxiliary inputs and as weights in the training loss function to emphasize clinically relevant tissue regions. Our approach achieved a 1.5\% and 1.4\% performance gain over current state-of-the-art methods, achieving 96.1 \% and 93.0\%, on the BMIRDS-LUAD and WSSS4LUAD datasets, respectively, demonstrating improved robustness in subtype classification and potential for integration into computational pathology workflows.

Track: 3. Imaging Informatics

Submission Number: 388