An optimized deep focused U-Net model for image segmentation

Neural network-based segmentation methods are an important advancement in medical image analysis. Issues with class imbalance pose a significant challenge in medical segmentation, with lesions often occupying a considerably smaller volume (hard objects) relative to the background (easy objects). Medical image segmentation tasks require the model to pay attention on specific parts of the image rather than the entire image. Models based on CNN and its variants like U-Net have shown promising results in this regard. However, they often suffer from unstable gradient during the training process. We propose an optimized deep focused U-Net (DF U-Net) model along with a novel learnable optimized focal loss (LOF) function. The DF U-Net is integrated with a novel block that comprises of five subblocks: efficient channel attention, squeeze and excitation, stochastic depth, residual block and dense block, collectively known as (ESSRD) in the decoder layers which helps to address the unstable gradient and overfitting issues. The proposed LOF loss optimizes itself during the training of the model and converges to optimal values within few epochs. We evaluated DF U-Net using four medical imaging data sets: DRIVE, BUS2017, CVC-Clinic and Kvasir-SEG. During the training process, DF U-Net significantly reduced the convergence time by sharply minimizing the loss to 0.001% and achieved an accuracy of 99.5%. The experimental results show that DF U-Net helps to mitigate the unstable gradient issue while focusing on both easy and hard objects equally.