Comparative Analysis of U-Net with Transfer Learning and Attention Mechanism for Enhanced Medical Image Segmentation

Medical image segmentation is an important component in computer-aided diagnosis and treatment. In this work, we study the application of the U-Net architecture to dermoscopy image segmentation, the former being a crucial stage in any algorithm attempting to detect and classify skin lesions. To enhance the segmentation task, we address two different approaches: we propose a novel method by integrating Attention gates inside the U-Net architecture (Attention U-Net), and we perform Transfer Learning for the encoder part of the U-Net using an array of pre-trained models leveraging ISIC-2017 dataset. The experimental configuration consists in a comparison between the common U-Net, the Attention U-Net and five Transfer Learning adaptations that each one uses one pre-trained model (VGG16, VGG19, MobileNetV2, ResNet50 and EfficientNetB0) in the encoder. The importance of Transfer Learning is highlighted through a thorough evaluation based on three key metrics: Loss, Intersection over Union (IoU), and Dice Similarity Coefficient (DSC). The DSC scores demonstrate superior segmentation performance of the enhanced architectures across benchmark analyses compared to the baseline U-Net model. Compared to other Transfer Learning variants, U-Net+MobileNetV2 is the overall best performer with a high DSC of 89.75%. It outperforms the respect DSC values of U-Net+ResNet50 (85.15%), U-Net+VGG19 (84.56%), U-Net+VGG16 (83.71%), and U-Net+EfficientNetB0 (50.44%). Additionally, due to the Attention Mechanism added to U-Net architecture, the DSC increases to 52.46%. These results collectively underline the efficacy of Transfer Learning, particularly with MobileNetV2, in substantially improving segmentation accuracy, as measured by the DSC metric.