A Supervised Lightweight and Efficient U-network for Skin Lesion Segmentation

Complex and computationally intensive models excel in segmentation tasks. However, they are impractical for resource-limited medical devices with restricted computing power. As a result, lightweight models have been developed as a solution, demonstrating promising performance. Yet, due to their architectural design and the diverse sizes and structures of skin lesions, these models struggle with inaccurate segmentation caused by feature loss during the feature extraction process. To unravel the improper segmentation, we propose a lightweight method that improves segmentation performance. Explicitly, our model is a U-network that applies various attention blocks. The network comprises spatial and channel attention to facilitate our model in focusing on the key features and innovative dilated spatial efficient attention to increase the balanced efficiency and performance of the model. In addition, an innovative pooling to improve feature retention and local variation reduction, and a channel attention skip connection to connect encoder-decoder features. We incorporate these modules and produce an efficient lightweight skin lesion segmentation network. Compared to standard UNet, our proposed model reduced more parameters by 29x, computational complexity by 17x and 2.64% and 1.44% improvement in mean Intersection over Union and Dice coefficient, respectively. Furthermore, we experimentally compare the performance of our method with other lightweight and large methods in the ISIC2017 and ISIC2018 skin lesion datasets. The results demonstrate that our proposed model attains improved performance with a 1.88% mean Intersection over Union and 0.65% Dice coefficient improvement compared to a LeaNet lightweight model.