Deep Learning-Based Methods for Automatic Diagnosis of Skin Lesions.

Area of Science:

• Dermatology
• Artificial Intelligence
• Medical Imaging

Background:

• Accurate diagnosis of skin lesions is crucial for effective treatment.
• Deep learning methods show promise in medical image analysis.
• Existing diagnostic systems may lack comprehensive accuracy.

Purpose of the Study:

• To develop a high-accuracy system for diagnosing skin lesions using deep learning.
• To propose a novel decision system integrating multiple classifiers.
• To enhance diagnostic decision-making through weighted fusion of classifier results.

Main Methods:

• Developed a neural network (NN) for melanoma vs. benign nevus differentiation.
• Employed pre-trained convolutional neural networks (CNNs) like GoogleNet, ResNet-101, and NasNet-Large with transfer learning.
• Utilized a classical image object detection method with feature extraction and support vector machine (SVM) classification.
• Integrated individual method results into a global fusion-based decision system with adaptive weighting.

Main Results:

• Evaluated NN performance using accuracy, specificity, sensitivity, and Dice coefficient.
• Determined classification accuracies for fine-tuned CNN architectures.
• Assessed the performance of the feature extraction and SVM method.
• Demonstrated superior accuracy of the proposed fusion-based system on two public datasets.

Conclusions:

• The proposed fusion-based decision system achieves high accuracy in skin lesion diagnosis.
• Integrating multiple deep learning and feature-based methods enhances diagnostic performance.
• The weighted fusion approach effectively leverages individual classifier strengths for improved decision-making.