MFA-Net: multi-scale feature aggregation network with background-aware module for ultrasound segmentation of thyroid nodules.

Background: The size and morphology of thyroid nodules are the essential basis for distinguishing benign and malignant in clinical diagnosis. However, achieving precise segmentation of these nodules in ultrasound images remains a significant task due to the weak and indistinct edges, low contrast, and complex internal structure. To tackle these challenges, our goal is to develop a multi-scale feature aggregation network (MFA-Net) with background-aware module (BAM), which can effectively and robustly segment ultrasound images of thyroid nodules. Methods: In MFA-Net framework, through the multi-scale feature aggregation module (MFAM), it effectively captures multi-scale context information and thus improves fine-grained details and global structure representation. Additionally, the BAM inhibits background noise, which allows for effective differentiation between nodules and surrounding tissue. To refine the segmentation performance, we add spatial and channel attentions to the residual decoder module (RDM). Results: The quantitative evaluation results showed that on the thyroid nodule 3493 (TN3K) dataset, MFA-Net achieved Dice of 0.8616, intersection over union (IoU) of 0.7586, accuracy of 0.9698 and Matthews correlation coefficient (Mcc) of 0.8457. On the thyroid gland 3583 (TG3K) dataset, the Dice, IoU, accuracy and Mcc reached 0.9857, 0.9718, 0.9977 and 0.9844. On the digital database thyroid image (DDTI) dataset, the Dice, IoU, accuracy and Mcc were 0.7483, 0.5981 0.9283 and 0.7078. On the BrainTumor dataset, MFA-Net obtained Dice of 0.8485, IoU of 0.7421, accuracy of 0.9949 and Mcc of 0.8469. Conclusions: These results outperform current leading models and confirm significant performance improvements. In addition, the MFAM, BAM and RDM demonstrate their robustness and adaptability in different segmentation scenarios when conducting independent ablation experiments.