Fusing fine-tuned deep features for skin lesion classification.

Area of Science:

• Dermatology
• Medical Imaging
• Artificial Intelligence

Background:

• Malignant melanoma is an aggressive skin cancer where early detection is critical for survival.
• Accurate differentiation of malignant skin lesions from benign ones is essential for effective diagnosis.
• Computerized classification of skin lesion images can significantly aid dermatologists in diagnosis.

Purpose of the Study:

• To develop a fully automatic computerized method for classifying skin lesions from dermoscopic images.
• To enhance diagnostic accuracy by effectively discriminating between malignant and benign skin lesions.
• To propose a novel ensemble scheme for convolutional neural networks (CNNs) for improved classification performance.

Main Methods:

• A novel ensemble scheme combining intra-architecture and inter-architecture CNN fusion was developed.
• Multiple sets of CNNs with varying architectures and fine-tuning settings were utilized.
• Deep features from each CNN were used to train support vector machine classifiers, with final predictions fused via averaging.

Main Results:

• The algorithm achieved an area under the receiver operating characteristic curve (AUC) of 87.3% for melanoma classification.
• An AUC of 95.5% was obtained for seborrheic keratosis classification on the ISIC 2017 dataset.
• The proposed method outperformed top-ranked algorithms in the challenge while maintaining simplicity.

Conclusions:

• The developed approach offers a reliable and robust method for feature extraction, model fusion, and classification of dermoscopic skin lesion images.
• This automated system shows significant potential to support clinical diagnosis and improve patient outcomes.
• The CNN ensemble scheme provides a powerful tool for analyzing complex medical image data.