A Study of Ultrasound Image Segmentation of Thyroid Nodules Based on Feature Adaptive Extraction and Gated Fusion Mechanisms

We use the context-gated feature attention module to filter multi-layer redundant features through the gating mechanism, and enhance cross-layer feature associations with the help of self-attention to suppress noise interference.Results The TN3K dataset is used in the study, and the results show that the accuracy of FAGF-Net reaches 95.59%, which is 3.77% higher than that of the baseline model UNet; the Dice coefficient is 90.60%, which is 3.55% higher.

The ablation experiments show that FC-Encoder, MFSE-Module and CGFA-Module improve the Dice coefficient by 1.40%, 0.69% and 2.41%, respectively, which verifies the necessity of each module. The visualisation results show that FAGF-Net significantly outperforms the comparison models in terms of boundary fitting smoothness, accuracy of tiny lesion localisation, and effectively reduces the phenomena of under-segmentation and over-segmentation.Conclusion FAGF-Net significantly improves the segmentation performance of ultrasound images of thyroid nodules through the innovative design of feature coupling and gating mechanism, providing a new solution for clinical automated diagnosis.