Improving the Diagnosis of Skin Biopsies Using Tissue Segmentation.

Area of Science:

• Digital pathology
• Computational oncology
• Artificial intelligence in healthcare

Background:

• Invasive melanoma is a deadly skin cancer requiring accurate pathological assessment.
• Pathologist interpretation of melanocytic lesions faces significant inter- and intra-observer variability.
• Machine learning (ML) shows promise in improving diagnostic accuracy in healthcare.

Purpose of the Study:

• To investigate the impact of semantic segmentation of tissue structures on the diagnostic performance of ML models for skin biopsy images.
• To assess if incorporating segmentation masks of epidermal and dermal nests improves melanoma diagnosis compared to using whole slide images alone.

Main Methods:

• Utilized whole slide images of skin biopsies for training and testing ML diagnostic models.
• Incorporated semantic segmentation masks of key tissue structures, specifically epidermal and cancerous dermal nests.
• Compared the performance of ML models trained with and without segmentation masks using F-score metric.

Main Results:

• A 6% improvement in F-score was achieved when segmentation masks were included in the ML pipeline.
• The inclusion of epidermal and cancerous dermal nest segmentation enhanced the diagnostic accuracy.
• Semantic segmentation proved beneficial for improving ML-based melanoma diagnosis.

Conclusions:

• Semantic segmentation of clinically important tissue structures can significantly improve ML-based diagnosis of invasive melanoma.
• This approach offers a potential solution to reduce observer variability in dermatopathology.
• Integrating AI with detailed image analysis holds promise for advancing skin cancer diagnostics.