Objective

2

Urinary tract stones in abdominal CT images present significant challenges for automatic segmentation， particularly due to their diverse morphologies， blurred boundaries， and low contrast with surrounding tissues. Traditional methods often struggle to achieve a balance of accuracy and robustness， particularly when detecting small objects and encountering complex background interference. Therefore， a DAST-UNet model is proposed that incorporates a dual attention mechanism to achieve high-precision automatic segmentation of urinary tract stones.

Methods

2

The proposed model utilizes the Swin-Unet backbone architecture and incorporates a CASI （channel and spatial interactive） module and a TDA （Token dependency attention） module to enhance the model’s feature representation capabilities from the channel-spatial dimension and the local-global semantic relationship level， respectively. A symmetric encoder-decoder architecture is employed to improve segmentation performance through multi-scale feature fusion and a progressive upsampling strategy. The CASI module enhances the model’s focus on the target region and effectively suppresses background noise， while the TDA module effectively captures fine-grained structural features and optimizes contextual information modeling by strengthening dependencies between Tokens.

Results

2

On a constructed urinary tract calculi CT image dataset， DAST-UNet outperforms classic models such as Res Unet， SegNet， and Swin-Unet across multiple key performance metrics， ultimately achieving a Dice coefficient of 80.34%， sensitivity of 80.21%， accuracy of 89.91%， and intersection over union （IoU） of 67.14% on the test set. The model loss function converges rapidly during training， with the validation set loss trend consistent with the training set， indicating that the network training process is stable and free of significant overfitting. Ablation experiments further validate the specific contributions of the CASI and TDA modules to performance improvement. The Swin-Unet + CASI model shows significant improvements in metrics such as Dice and sensitivity. The introduction of TDA further improves performance， resulting in DAST-UNet performing the best among all combinations.

Conclusions

2

The proposed DAST-UNet model achieves a coordinated modeling of local details and global context， effectively improving the accuracy of automatic segmentation of urinary tract calculi in CT images. The experimental results fully verify the positive effect of the dual attention mechanism on small target recognition and complex background processing， and provide a reliable and scalable automated solution for related clinical applications.