Comparative Study of Deep Learning Architectures for Skin Lesion Segmentation

Abstract

Early finding of skin cancer is very important because it can save the patient life if the problem is found early. Doctors usually check the skin images by hand, but this way takes a lot of time and the results can change from one doctor to another. So, using computer and deep learning can make this process faster and more accurate. In this paper, two models are used, U-Net and DeepLabV3+, to detect and segment the lesion area from the ISIC-2020 dataset. Each model was trained with two encoders, ResNet-50 and EfficientNet-b0, to see which one gives better results. The images were resized and augmented before training to help the models learn better. From the results, EfficientNet-b0 worked better than ResNet-50 and gave smoother training. The best accuracy was from U-Net with EfficientNet-b0, which reached Dice 0.920 and IoU 0.859. This work shows that deep learning can really help doctors to find skin cancer faster and more reliable