Skin Lesion Classification Using Collective Intelligence of Multiple Neural Networks

Area of Science:

• Dermatology and Artificial Intelligence
• Medical Imaging Analysis
• Computational Pathology

Background:

• Early detection of skin cancer is crucial for effective treatment and preventing metastasis.
• Accurate computer-aided diagnosis systems are needed to support clinicians in identifying malignant skin lesions.
• Deep learning shows promise for analyzing complex medical images like skin lesions.

Purpose of the Study:

• To develop and evaluate a novel skin lesion classification system utilizing deep learning and collective intelligence.
• To enhance the accuracy of early detection for seven types of skin lesions, including melanoma.
• To create an ensemble model that combines multiple convolutional neural networks for improved diagnostic performance.

Main Methods:

• A collective intelligence-based system was designed, integrating multiple convolutional neural networks (AlexNet, GoogLeNet, MobileNet-V2, Xception, ResNet, InceptionResNet-V2, DenseNet201).
• Networks were trained on the HAM10000 dataset for skin lesion classification.
• A weight matrix was derived from individual network performances to create a multi-network ensemble system with a collective decision block for fused predictions.

Main Results:

• The proposed collective intelligence system achieved a validation accuracy approximately 3% higher than the best-performing individual convolutional neural network.
• The ensemble approach demonstrated enhanced performance in classifying various skin lesions, including melanoma.
• Individual network performances were analyzed to assign weights for the ensemble decision-making process.

Conclusions:

• The developed collective intelligence system offers a more accurate approach to skin lesion classification compared to individual deep learning models.
• This system has the potential to significantly aid medical professionals in the early and accurate detection of skin cancer.
• Further validation and clinical integration of this AI-driven diagnostic tool are warranted.