A novel hybrid convolutional and transformer network for lymphoma classification.

Area of Science:

• Pathology
• Computer Science
• Artificial Intelligence

Background:

• Lymphoma classification from Whole Slide Images (WSIs) is challenging due to subtype morphological similarities.
• Traditional methods lack objectivity and consistency, necessitating advanced AI solutions.
• Current models struggle to integrate local and global features for accurate diagnosis.

Purpose of the Study:

• To develop and evaluate a hybrid deep learning framework for precise lymphoma subtype classification.
• To improve the interpretability of AI models in hematopathology.
• To address the limitations of existing models in capturing both local and global image features.

Main Methods:

• Proposed a Hybrid Convolutional and Transformer Network for Lymphoma Classification (HCTN-LC).
• Employed a dual-pathway architecture combining SqueezeNet for local features and Vision Transformer (ViT) for global context.
• Introduced a Feature Fusion and Enhancement Module (FFEM) for dynamic feature integration.

Main Results:

• HCTN-LC achieved high performance on a WSI dataset of CLL, FL, and MCL subtypes.
• Achieved an overall accuracy of 99.87%, sensitivity of 99.87%, specificity of 99.93%, and AUC of 0.9991.
• Grad-CAM visualizations confirmed the model's focus on diagnostically relevant regions, demonstrating interpretability.

Conclusions:

• The HCTN-LC framework offers superior performance and interpretability for lymphoma subtype classification.
• Demonstrates potential for clinical deployment in low-resource settings.
• Provides a robust AI tool for hematopathological diagnosis, improving patient care.