Weakly Supervised Classification of Mohs Surgical Sections Using Artificial Intelligence.

Area of Science:

• Dermatology
• Artificial Intelligence
• Medical Imaging

Background:

• Basal cell carcinoma (BCC) incidence is rising, increasing healthcare burdens.
• Histopathologic diagnostics for Mohs micrographic surgery (MMS) are crucial for BCC treatment.
• Accurate BCC detection in MMS whole-slide images presents diagnostic challenges.

Purpose of the Study:

• To develop and evaluate a deep learning model for BCC detection in MMS images.
• To integrate weakly supervised learning with interpretable segmentation using attention maps.
• To enhance diagnostic accuracy and interpretability in BCC identification.

Main Methods:

• Developed a deep learning model combining weakly supervised learning and attention-based segmentation.
• Utilized datasets from two medical centers for training and internal testing.
• Validated the model on an independent external dataset without fine-tuning.

Main Results:

• Achieved an average AUC of 0.958 on internal testing and 0.934 on an external dataset.
• Attention maps provided insights into model decision-making and highlighted critical regions.
• Tumor localization sensitivity was 0.853, with an average of 8 false positives per slide, reduced to 2 false positives per slide with 0.873 sensitivity when filtering small detections.

Conclusions:

• The deep learning model demonstrates high effectiveness and robustness in detecting BCC in MMS images.
• Attention maps enhance model interpretability, aiding dermatopathologists and surgeons.
• This AI tool shows promise for improving BCC diagnosis and MMS procedures.