Automatic calculation method for stenosis ratio based on dialysis access ultrasound image segmentation

Background

Dialysis Access (DA) stenosis impacts hemodialysis efficiency and patient health, necessitating exams for early lesion detection. Ultrasound is widely used due to its non-invasive, cost-effective nature. Assessing all patients in large hemodialysis facilities strains resources and relies on operator expertise. Furthermore, it heavily relies on the experience and expertise of the operator. Therefore, it is essential to explore methods for the automatic analysis of DA ultrasound images to accurately calculate the stenosis ratios, thereby enhancing both diagnostic accuracy and treatment efficiency.

Purpose

This study is aimed at employing image segmentation networks to conduct precise segmentation of the ultrasound images of DA lumens and automatically classify the types of stenosis. The segmentation outcomes are processed by means of morphological processing techniques for the automatic calculation of the DA stenosis ratio, thus enhancing the daily diagnostic efficiency of physicians and providing a substantial quantitative foundation for clinical decision-making.

Methods

Firstly, our study introduces a deep neural network-based approach for vascular lumen segmentation and classification, termed Vessel Lumen Segmentation and Classification-Net (VLSC-Net), aimed at the precise segmentation of the DA lumen in ultrasound images. We conducted comparative analyses of our network against U-Net, TransUNet, MultiResUnet, and ResUNet using metrics such as mean Intersection over Union (mIoU), Dice score, Accuracy, Hausdorff Distance (HD), and Average Symmetric Surface Distance (ASSD). A five-fold cross-validation was performed on a dataset comprising 1710 images for both comparison experiments and ablation studies; specifically, the training set included 1368 images while the test set contained 342 images. The significance of observed differences was assessed using the Mann-Whitney U-test. To prevent the increase in the chance of making a Type I error (false positive) that occurs when many simultaneous tests are being conducted, we used the Bonferroni correction to address the problem of multiple comparisons. Since we did four groups of comparisons, the significance level ($\alpha$) is adjusted by dividing it by 4. Secondly, we utilized morphological processing alongside feature extraction techniques to accurately delineate the edges of the lumen. This facilitated precise measurements of critical stenosis segment parameters. Finally, we automatically calculated the Long-axis Diameter Stenosis Ratio (LDSR) and Short-axis Area Stenosis Ratio (SASR) utilizing methods from the European Carotid Surgery Trial based on parameters derived from these calculations.

Results

VLSC-Net demonstrated superior performance compared with traditional segmentation methods, effectively handling image artifacts while maintaining a compact structure. The mIoU, Dice score, Accuracy, HD, and ASSD were 0.9563, 0.9777, 0.9976, 4.542, and 0.460, respectively, and showed significant differences from the results of U-Net (p 0.0125). An evaluation involving 1710 images from 62 patients indicated that our method delivers high-precision and reliable stenosis ratio and classification outcomes within an average processing time of 164 ms. Furthermore, the average errors for LDSR and SASR were found to be 1.4% and 7.8%, respectively.

Conclusions

Our approach greatly enhances diagnostic efficiency for medical personnel, offering reliable and objective evidence for clinical assessment and decision-making in DA stenosis treatment, thereby reducing the risk of complications associated with DA stenosis.