A Novel Instance Segmentation Method for Real-Time Detection of Thyroid Nodules in Ultrasound Videos: A Multi-Center Study.

OBJECTIVES

Thyroid nodules are one of the most common thyroid disorders and can be categorized into benign and malignant thyroid nodules. Currently, the initial diagnosis is made clinically by ultrasonography, and the determination of thyroid nodules is highly dependent on the operating experience and maneuvers of the physician, resulting in fluctuating diagnostic accuracy. Recent studies have shown that using deep learning-based automated diagnostic tools can assist in the segmentation of individual nodules in thyroid ultrasound images. However, in reality, a single frame typically contains multiple nodules.

METHODS

Based on the above, we propose a segmentation model for multiple instances of ultrasound videos to improve the accuracy and detection rate of thyroid nodules. Firstly, we introduced a variable convolutional network in the backbone network part to improve the model's ability to extract nodule texture features. Secondly, we proposed a bidirectional mamba module to enhance the model's long-range modeling capability for video data and improve the accuracy, and a Sobel edge operator module to improve the accuracy of nodal boundaries. Finally, we proposed a detection rate metric to evaluate the model's efficiency, and thus assessed the model's performance in the clinic.

RESULTS

Results on a validation set from multiple hospitals showed that our model performed well, with a 15.1% improvement in dice score and a 19.3% improvement in detection rate to 89.5% over the existing techniques.

CONCLUSION

The multi-nodule instance segmentation model for thyroid ultrasound video we proposed has significantly improved the accuracy and effectiveness of thyroid nodule segmentation.