A Novel Multi-Task Learning Network Based on Melanoma Segmentation and Classification with Skin Lesion Images

Area of Science:

• Dermatology
• Medical Imaging
• Artificial Intelligence

Background:

• Melanoma is a dangerous skin cancer with high mortality.
• Early detection is crucial for improving survival rates.
• Automatic detection systems can aid in early diagnosis.

Purpose of the Study:

• To develop and evaluate a multi-task learning approach for melanoma recognition using dermoscopy images.
• To enhance the accuracy and efficiency of automatic melanoma detection.

Main Methods:

• Image pre-processing using max pooling, contrast, and shape filters.
• Lesion segmentation utilizing a VGGNet model-based FCN Layer architecture.
• Deep learning classification with pre-trained convolutional neural networks.

Main Results:

• High performance in lesion segmentation (96.99% accuracy, 98.41% sensitivity).
• Excellent performance in melanoma classification (97.73% accuracy, 95.67% sensitivity).
• Demonstrated effectiveness of the deep learning approach on the ISIC dataset.

Conclusions:

• The proposed multi-task learning approach is effective for automated melanoma detection and classification.
• The system shows potential for improving early diagnosis and patient outcomes.
• Deep learning models offer a promising avenue for advancing dermatological diagnostics.