Semantic Segmentation for Pixel-Wise Visualization of Breast Cancer in Deep Ultraviolet-Excited Fluorescence Images

Area of Science:

• Pathology
• Medical Imaging
• Artificial Intelligence

Background:

• Microscopy with Ultraviolet Surface Excitation (MUSE) provides rapid fluorescence imaging of tissue surfaces.
• MUSE images present distinct visual characteristics compared to conventional hematoxylin and eosin (H&E) staining, posing challenges for accurate cancer delineation in routine pathology.

Purpose of the Study:

• To investigate the feasibility of deep learning-based semantic segmentation for pixel-wise breast cancer detection in MUSE images.
• To evaluate the performance of different deep learning models in identifying cancerous regions within MUSE images.

Main Methods:

• Fresh breast tissues from 30 mastectomy patients were stained and imaged using MUSE.
• A dataset of 150 cancerous and 300 non-cancerous MUSE images was manually annotated.
• Deep learning models, including a cancer-only (CO) and a cancer plus non-cancer (CN) model, were trained and evaluated using five-fold nested cross-validation.
• Sliding window-based majority voting was employed as a post-processing technique.

Main Results:

• The cancer-only (CO) model achieved a higher Dice score (0.7478) than the cancer plus non-cancer (CN) model (0.7343).
• Post-processing with sliding window majority voting improved Dice scores for both models (CO: 0.7984, CN: 0.7849).
• The results demonstrate the effectiveness of semantic segmentation in reducing false positives and enhancing cancer visualization.

Conclusions:

• Deep learning-based semantic segmentation is feasible for visualizing breast cancer regions in MUSE images.
• This approach provides a foundation for future quantitative analyses using MUSE imaging in pathology.
• The study highlights the potential of AI in improving cancer detection and analysis from novel imaging modalities.