SERAU-Net: Snapshot Ensemble-Aided Region-Aware Attention-Based U-Net for Skin Lesion Segmentation

Skin images are segmented by separating them into discrete areas to emphasize particular structures, such as cells. This procedure helps dermatologists develop automated diagnosis systems to focus on the required segments, improving the accuracy and productivity of their analyses. This study proposes a new model, called SERAU-Net (Snapshot Ensemble-aided RegionAware Attention-based U-Net) for skin lesion segmentation. A U-Net architecture with the backbone of RA-Net is used, which incorporates a region-aware attention mechanism along with depth-wise separable convolutions. While the depth-wise separable convolutions are utilized to reduce trainable parameters, the region-aware attention mechanism is employed to refine feature information for pixel-by-pixel classification. The snapshot ensemble is a technique for improving the performance of a convolutional neural network (CNN), which entails training a CNN over a number of iterations to produce multiple learned models (referred to as snapshots). The final decision is subsequently made by combining these snapshots, which lessens the computing load. We have used this method in this study to generate five RANet models' snapshots. We have employed the Cosine Annealing learning rate schedule for this purpose. To evaluate our model, we have used two datasets, PH2 and ISIC 2016, for image segmentation related to skin cancer. Our model has produced results that are on par with some recent models. The detailed results are available at: SERAUNET-Detailed-Results.