Breast Cancer Classification in Deep Ultraviolet Fluorescence Images Using a Patch-Level Vision Transformer...

Area of Science:

• Oncology
• Medical Imaging
• Artificial Intelligence

Background:

• Breast-conserving surgery (BCS) requires precise intraoperative margin assessment to balance cancer removal and tissue preservation.
• Deep ultraviolet fluorescence scanning microscopy (DUV-FSM) offers rapid whole surface imaging (WSI) for excised breast tissue, differentiating malignant from normal tissue.
• Classifying breast cancer from high-resolution DUV WSIs presents challenges due to complex histopathological features.

Purpose of the Study:

• To develop and evaluate a DUV WSI classification framework for accurate breast cancer margin assessment.
• To enhance the interpretability and diagnostic accuracy of DUV WSI analysis for intraoperative margin assessment.

Main Methods:

• A patch-level vision transformer (ViT) model was employed to analyze DUV WSIs, capturing both local and global histopathological features.
• Grad-CAM++ saliency weighting was integrated to highlight critical spatial regions within the WSIs, improving model interpretability.
• A 5-fold cross-validation strategy was utilized to rigorously assess the performance of the proposed classification framework.

Main Results:

• The proposed DUV WSI classification framework achieved a high classification accuracy of 98.33% for distinguishing benign from malignant breast tissue.
• The ViT model effectively captured complex histopathological features from DUV WSIs, outperforming conventional deep learning methods.
• Grad-CAM++ saliency weighting enhanced the interpretability of the model's predictions, identifying diagnostically relevant tissue areas.

Conclusions:

• The developed DUV WSI classification framework, utilizing a ViT model and Grad-CAM++, significantly improves diagnostic accuracy for intraoperative margin assessment in BCS.
• This approach offers a promising tool for real-time, accurate classification of breast tissue during surgery, potentially reducing re-excision rates.
• The framework demonstrates the potential of advanced AI techniques in enhancing histopathological analysis of DUV fluorescence images for breast cancer management.