Modified SQEU-Net: An Enhanced U-Net Architecture With Federated Learning For Urolithiasis Segmentation And Type Classification

Abstract

The precise segmentation of Urolithiasis on the basis of the computed tomography (CT) images is a key requirement to consistent diagnosis, treatment planning, and quantitative evaluation in urology. The large range of variability in stone size, shape, and distribution of intensities and the existence of the surrounding structures of the anatomy create serious problems in automated segmentation procedures, though. In order to overcome these shortcomings, this paper has suggested a new improved version of U-Net, SQEU-Net, a deep learning network, especially tailored to kidney stone segmentation. The proposed model incorporates dilated convolutions in the encoder to allow multi-scale contextualization without much spatial decrease, residual learning to enhance its ability to optimize the deep networks, and squeeze-and-excitation (SE) to recalibrate channel-wise feature responses in a dynamic manner. Besides, attention-gated skip connections are used to selectively pass clinically relevant features between the encoder and the decoder, eliminating background noise and irrelevant anatomical structures. Using the CT kidney stone data, the model is trained and tested on the data with standard segmentation measures of Dice Similarity Coefficient (DSC), Intersection over Union (IoU), accuracy, precision, recall, and F1-score measures. Comparative experiments with the state-of-the-art in the recent past show that SQEU-Net always performs better in terms of segmentation especially with regard to defining small and irregularly shaped stones. The findings confirm the usefulness of integrating contextual features learning, channel-wise attention, and spatial attention systems to achieve powerful kidney stone segmentation, one of the roles that SQEU-Net can play is a clinical reliable decision-support system.