Convolutional Neural Network Approach to Classify Skin Lesions Using Reflectance Confocal Microscopy.

Area of Science:

• Dermatology
• Medical Imaging
• Artificial Intelligence

Background:

• Skin cancers, including melanoma, have high morbidity and mortality rates.
• Early and accurate diagnosis is crucial for effective treatment.
• Reflectance confocal microscopy (RCM) offers non-invasive virtual biopsies but requires specialized expertise.

Purpose of the Study:

• To develop an AI-based tool to assist in the classification of skin lesions using RCM mosaics.
• To improve the accuracy and accessibility of early skin cancer diagnosis.

Main Methods:

• A convolutional neural network, based on the ResNet architecture, was developed.
• The network was trained on a dataset of 429 RCM mosaics.
• The dataset included three classes: melanoma, basal cell carcinoma, and benign naevi, with histopathological confirmation.

Main Results:

• The classification system achieved an accuracy of 87% on the test set.
• The AI model's accuracy surpassed that of experienced medical confocal users.
• The system demonstrated potential for aiding in the early detection of melanoma.

Conclusions:

• The proposed convolutional neural network is a viable tool for classifying skin lesions from RCM images.
• This AI approach can support clinicians in non-invasive, early detection of skin cancers.
• Further development could enhance diagnostic capabilities in dermatology.