Cross-platform multi-cancer histopathology classification using local-window vision transformers.

Area of Science:

• Digital Pathology
• Artificial Intelligence in Oncology
• Computational Biology

Background:

• Accurate histopathological image classification is crucial for cancer diagnosis and treatment.
• Current deep learning models often lack generalizability and transparency for clinical applications.
• Existing models are typically limited to single-cancer classification.

Purpose of the Study:

• To develop a unified deep learning framework, CancerDet-Net, for multi-cancer histopathological image classification.
• To enhance model generalizability, interpretability, and clinical applicability.
• To achieve high accuracy in classifying diverse cancer subtypes.

Main Methods:

• Integration of separable convolutional layers, Vision Transformer (ViT) blocks with local-window sparse self-attention.
• Application of a Hierarchical Multi-Scale Gated Attention Mechanism (HMSGA) combined through Cross-Scale Feature (CSF) Fusion.
• Development of explainable AI (XAI) visualizations for model transparency.

Main Results:

• CancerDet-Net achieved a top-performing accuracy of 98.51% in classifying nine histopathological subtypes across four major cancer types.
• The model demonstrated strong generalizability across datasets.
• Integrated XAI visualizations provided transparency for clinical interpretation.

Conclusions:

• CancerDet-Net offers a comprehensive and unified framework for multi-cancer classification in digital pathology.
• The model's high accuracy, generalizability, and interpretability represent a significant advancement.
• Deployment readiness via web and mobile platforms facilitates real-time clinical use, enhancing AI-driven cancer diagnosis.