Modified U-NET Architecture for Segmentation of Skin Lesion

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• Accurate skin lesion segmentation is crucial for reliable classification of dermoscopic images.
• Existing segmentation algorithms struggle with the inherent complexities of skin lesions, such as fuzzy borders and class variances.
• Improved segmentation accuracy is needed to meet industry standards in dermatological diagnostics.

Purpose of the Study:

• To propose a modified U-Net architecture for accurate and automatic segmentation of skin lesions in dermoscopic images.
• To enhance nodule extraction precision by adjusting feature map dimensions and increasing the number of kernels.
• To evaluate the model's performance using various hyperparameters and data augmentation techniques.

Main Methods:

• A modified U-Net architecture was developed, featuring adjusted feature map dimensions and increased kernels for improved feature extraction.
• The model was trained and evaluated on the PH2 dataset, incorporating data augmentation techniques.
• Hyperparameter tuning included optimizing epochs, batch size, and optimizer type (Adam).

Main Results:

• The proposed modified U-Net architecture demonstrated enhanced accuracy in segmenting skin lesions from dermoscopic images.
• The optimal performance was achieved using the Adam optimizer with a batch size of 8 and 75 epochs.
• Data augmentation and architectural modifications contributed to more precise nodule extraction.

Conclusions:

• The modified U-Net architecture offers a promising solution for accurate and automatic skin lesion segmentation in dermoscopic images.
• The study highlights the importance of architectural modifications and hyperparameter optimization for improving segmentation performance.
• This approach has the potential to advance computer-aided diagnosis in dermatology.