Dermoscopic Image Classification Method Using an Ensemble of Fine-Tuned Convolutional Neural Networks.

Area of Science:

• Dermatology
• Medical Imaging
• Computer Science

Background:

• Dermoscopic image analysis faces challenges like large intra-class variations, subtle inter-class differences, low contrast, and limited, imbalanced datasets.
• Accurate classification of skin lesions is crucial for early diagnosis and treatment of skin cancer.

Purpose of the Study:

• To develop an effective deep learning model for classifying dermoscopic images to identify malignant skin lesions.
• To address the limitations of existing datasets and improve classification performance.

Main Methods:

• An ensemble learning strategy was employed, integrating three pre-trained convolutional neural network (CNN) models: Xception, ResNet50, and Vgg-16.
• Transfer learning and fine-tuning were performed on the ISIC 2016 Challenge skin dataset.
• A weighted fusion ensemble strategy was used to combine the predictions of the base models.

Main Results:

• The proposed ensemble model achieved an accuracy of 86.91%, precision of 85.67%, recall of 84.03%, and F1-score of 84.84%.
• These performance metrics surpassed those of the individual base models and some classical methods.
• The method demonstrated effectiveness in classifying dermoscopic images, particularly in distinguishing malignancy.

Conclusions:

• The ensemble of fine-tuned CNNs provides a robust and effective method for dermoscopic image classification.
• This approach successfully overcomes common challenges in skin lesion image analysis, offering improved diagnostic potential.
• The study validates the feasibility and superiority of the proposed ensemble strategy for skin cancer detection.