UMAM-NET: ultrasound thyroid nodule malignancy grading network based on multi-subnet attention mechanism

• Published in: Pattern analysis and applications : PAA Vol. 28; no. 1
• Main Authors: Bi, Hui, Wang, Fan, Xiong, YuHao, Dong, ZhaoHui, Jiang, Yibo, Zhao, Tong, Zheng, Yineng
• Format: Journal Article
• Language: English
• Published: London Springer London 01.03.2025; Springer Nature B.V
• Subjects: Accuracy; Artificial neural networks; Computer Science; Datasets; Feature extraction; Image resolution; Ionizing radiation; Machine learning; Modules; Nodules; Original Article; Pattern Recognition; Recall; Texture; Thyroid gland; Ultrasonic imaging; Attention mechanism; Multi-subnet; Ultrasound; Thyroid
• ISSN: 1433-7541; 1433-755X
• DOI: 10.1007/s10044-024-01404-7

Background and Objectives: Thyroid nodules are one of the most common thyroid diseases, and their incidence has been on the rise in recent years. Ultrasound imaging, due to its low cost and no ionizing radiation, has become the preferred method for imaging thyroid nodules. Accurate assessment of the malignancy grade of thyroid nodules is crucial to ensure the accuracy of subsequent examination and treatment. Texture and shape are key features for determining the nature of thyroid nodules. Despite the excellent performance of convolutional neural networks (CNNs) in image feature extraction and aggregation, the low resolution and high noise characteristics of ultrasound images still pose challenges for existing CNN models in identifying texture and shape. Methods: To address this challenge, we propose a thyroid nodule malignancy grading network based on a multi-subnet attention mechanism (UMAM-NET). In the feature extraction stage, we innovatively introduce the multi-subnet attention module. The module designs two parallel subnets, aiming to enhance the model’s attention to the texture and shape of thyroid nodules. Results: Compared to other deep learning models, the proposed UMAM-NET performs better in the malignant grading task of thyroid nodules. It demonstrates excellent performance on public datasets, achieving the best results in Recall (93.1%), F1-score (95.4%), and Accuracy (98.4%). Similarly, it also shows outstanding performance on the sub-collected dataset, with Recall (91.8%), F1-score (92.0%), and Accuracy (94.4%). Conclusion: Our proposed UMAM-NET, based on multi-subnet attention mechanism, provides a promising approach for accurate assessment of thyroid nodule malignancy grade, which can be of great value in clinical practice.