Deep Learning Approach to Classify Cutaneous Melanoma in a Whole Slide Image

Area of Science:

• Dermatopathology
• Computational pathology
• Artificial intelligence in medicine

Background:

• Histopathological diagnosis of cutaneous melanocytic lesions, while generally accurate, faces challenges with certain cases like melanoma.
• Diagnostic controversies arise with melanoma mimicking benign nevi, amelanotic features, or in situ presentations.
• Clinical-pathological correlation is the established gold standard for melanoma diagnosis.

Purpose of the Study:

• To investigate the feasibility of applying deep learning algorithms for the automated classification of cutaneous melanoma.
• To develop and evaluate AI models capable of assisting surgical pathologists in diagnosing melanocytic lesions from whole-slide images (WSIs).

Main Methods:

• A dataset of 66 whole-slide images (WSIs), comprising 33 melanomas and 33 non-melanomas, was utilized for model training.
• Weakly supervised learning techniques were employed to train the deep learning models.
• Model performance was assessed on an independent test set of 90 WSIs (40 melanomas, 50 non-melanomas).

Main Results:

• The best performing deep learning model achieved an Area Under the Receiver Operating Characteristic Curve (ROC-AUC) of 0.821 at the whole-slide image (WSI) level.
• At the tile level, the model demonstrated a higher ROC-AUC of 0.936, indicating strong performance in identifying melanoma at a finer resolution.
• These preliminary results suggest significant potential for AI in melanoma classification.

Conclusions:

• Deep learning models can effectively classify cutaneous melanoma in whole-slide images (WSIs).
• The developed AI approach shows potential as a valuable tool to support surgical pathologists, particularly in complex or time-consuming cases.
• Further research and validation are warranted to integrate this technology into routine diagnostic workflows.