Automatic segmentation of tophi in ultrasound images based on deep learning: a feasibility study

Abstract Background Tophi are crucial for the assessment and treatment monitoring of gout. Musculoskeletal ultrasound serves as an important complementary tool to physical examination of tophi. However, its clinical utility is limited by operator dependency, the need for tedious manual delineation, and the difficulty in identifying tophi with atypical sonographic features. This study aims to explore the feasibility of deep learning in the automatic segmentation of tophi in ultrasound images by evaluating different segmentation algorithms. Methods A total of 730 original tophi ultrasound images were collected and randomly divided into a training set (558 images), a validation set (89 images) and an internal test set (83 images) at the patient level. An independent external test set containing 30 images was additionally included for generalization assessment. Four deep neural networks—U-Net, Attention U-Net, TransUNet, and Deeplabv3+—were applied for tophus segmentation in this study. The Jaccard Similarity Coefficient (JSC) and Dice Similarity Coefficient (DSC) were adopted as the evaluation metrics on the test set. The number of parameters and per-image inference time of each model were also compared. Segmentation performance of all models was further compared with annotations from junior and senior sonographers, and false positive rates were calculated on non-tophaceous gout ultrasound images to verify model specificity. Results All four models exhibited good convergence during the training and validation process. In the internal test set, Deeplabv3 + achieved the best performance in terms of average JSC (64.60%) and average DSC (76.77%), followed closely by TransUNet (average JSC 63.59%, average DSC 76.29%); U-Net had the smallest number of parameters (31.00 million), along with the fastest per-image inference time (68.78 ms). Model performance declined slightly on the external test set, with no statistically significant inter-model differences observed. The segmentation accuracy of Deeplabv3 + and TransUNet was comparable to senior physicians and superior to junior physicians. When applied to images without tophi, all models generated only very small segmented regions with overall low false positive rates, and TransUNet showed the lowest false positive rate across most non-tophaceous lesions. Conclusions This study demonstrates the strong feasibility of deep learning for automated tophus segmentation in ultrasound images, laying a foundation for quantitative assessment of tophus volume and enabling the development of standardized, objective tools to support gout assessment and treatment monitoring.