Deep Learning for Automated Segmentation of Basal Cell Carcinoma on Mohs Micrographic Surgery Frozen Section Slides

Area of Science:

• Dermatopathology
• Computational Pathology
• Artificial Intelligence in Medicine

Background:

• Deep learning models have shown success in classifying basal cell carcinoma (BCC) from histopathologic images.
• Segmentation models are crucial for tumor localization in Mohs surgery (MMS) frozen sections but lack clinical utility.

Purpose of the Study:

• To develop and evaluate a segmentation model for localizing BCC on MMS frozen section slides.
• To assess the model's performance across different BCC subtypes.

Main Methods:

• Utilized 348 BCC patient frozen tissue slides for whole slide imaging.
• Manually annotated BCC foci using the Grand Challenge platform.
• Employed the Ultralytics YOLOv8 model for segmentation, with data split for training, validation, and testing.

Main Results:

• Achieved an overall sensitivity of 0.71 and specificity of 0.75 for BCC tumor localization.
• Performance varied by subtype: nodular BCC (sensitivity 0.87), superficial BCC (sensitivity 0.79), micronodular BCC (sensitivity 0.74), and morpheaform/infiltrative BCC (sensitivity 0.51).
• Specificity also showed subtype variations, with micronodular BCC achieving 0.83 and nodular BCC 0.59.

Conclusions:

• A deep learning segmentation model can localize BCC on MMS frozen sections with notable sensitivity and specificity.
• Model performance is dependent on the specific BCC subtype.
• Development of more accurate and clinically relevant performance metrics for segmentation studies is essential.