Deep Hybrid Convolutional Neural Network for Segmentation of Melanoma Skin Lesion

Area of Science:

• Medical image analysis
• Artificial intelligence in dermatology
• Computational pathology

Background:

• Melanoma diagnosis relies on manual segmentation of skin lesions, which is inefficient and prone to errors.
• Early detection of melanoma significantly improves patient prognosis and survival rates.
• Automated segmentation methods are needed to enhance the accuracy and efficiency of melanoma diagnosis.

Purpose of the Study:

• To develop and evaluate an improved automatic image segmentation algorithm for melanoma detection.
• To enhance the precision and speed of skin lesion segmentation compared to existing methods.

Main Methods:

• An EfficientUNet++ algorithm was developed, integrating EfficientNet with the U-Net architecture.
• The algorithm was trained and validated on the PH2 and International Skin Imaging Collaboration (ISIC) skin lesion datasets.
• Performance was compared against other common segmentation models using Dice coefficient, Intersection over Union (IoU), and loss value metrics.

Main Results:

• EfficientUNet++ demonstrated superior performance on the PH2 dataset, achieving a 93% Dice coefficient and 96% IoU.
• On the ISIC dataset, EfficientUNet++ achieved a 96% Dice coefficient and 94% IoU.
• The algorithm consistently showed lower loss values compared to other models across both datasets, indicating improved segmentation accuracy.

Conclusions:

• The EfficientUNet++ model offers an efficient and precise solution for automated skin lesion segmentation in melanoma detection.
• Integration of EfficientNet and residual units within the U-Net architecture significantly improves segmentation performance.
• This automated approach has the potential to aid clinicians in early and accurate melanoma diagnosis.