Skin Lesion Segmentation Using U-Net With Different Backbones: Comparative Study

Skin lesion segmentation is a critical step in computer-aided diagnosis systems for dermatology, aiding in the early detection of skin diseases. In this research, we present a robust approach for skin lesion segmentation utilizing the PH2 dataset, which comprises 200 dermoscopic images along with their medical segmentations. The proposed method integrates the U-Net architecture using VGG16 as a backbone, a well-established convolutional neural network model, and employs data augmentation techniques to enhance the segmentation performance. The results of the study demonstrate exceptional segmentation accuracy and effectiveness. The quantitative evaluation metrics showcase the model's proficiency, with an accuracy of 0.9459, an Intersection over Union (IoU) score of 0.9483, a Dice coefficient of 0.8826, precision of 0.9353, and a recall of 0.9151. These remarkable metrics highlight the model's precision in outlining skin lesions within dermoscopic images. This research contributes to the field of dermatology and medical imaging by providing a reliable and efficient approach for skin lesion segmentation. The combination of the U-Net architecture with data augmentation techniques significantly improves the accuracy and robustness of skin lesion segmentation, which is vital for early disease detection and accurate diagnosis. This work has the potential to advance the development of computer-aided diagnostic systems in dermatology, ultimately improving patient outcomes and facilitating more efficient clinical workflows.