Deep convolutional neural network with fusion strategy for skin cancer recognition: model development and validation

Area of Science:

• Dermatology
• Artificial Intelligence
• Medical Imaging

Background:

• Skin cancer classification relies on accurate diagnosis from clinical images.
• Deep learning models show promise but often require large datasets.
• Developing efficient systems for smaller datasets is crucial for clinical application.

Purpose of the Study:

• To develop an accurate and efficient deep learning-based skin cancer classification system.
• To utilize model fusion and lifelong learning for improved performance on limited data.
• To evaluate the system's accuracy against traditional deep learning models and dermatologists.

Main Methods:

• A novel deep convolutional neural network, SkinFLNet, was developed.
• SkinFLNet incorporates model fusion and lifelong learning techniques.
• The system was trained on 1215 clinical skin tumor images and validated on 463 images.

Main Results:

• SkinFLNet achieved 85% overall accuracy, 82% recall, and 93% specificity.
• The system outperformed other deep convolutional neural network models.
• SkinFLNet's performance was comparable or superior to that of three board-certified dermatologists.

Conclusions:

• SkinFLNet offers an efficient skin cancer classification system trainable on small datasets.
• The model fusion and lifelong learning approach enhances classification accuracy.
• This technology has the potential to improve clinical skin cancer screening accuracy.