RTS-Net: thyroid nodule segmentation network integrating dual-path attention and graph convolution.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Thyroid ultrasound interpretation is subjective and inefficient due to image noise and operator dependence.
• Current deep learning models often neglect anatomical information, limiting performance on complex cases.

Purpose of the Study:

• To develop a novel deep learning segmentation network, RTS-Net, for improved thyroid nodule detection.
• To enhance feature representation and boundary integrity in thyroid ultrasound images.

Main Methods:

• RTS-Net utilizes a dual-path attention mechanism (spatial and channel) and a cascaded graph convolution decoding architecture.
• Multi-scale feature pyramid fusion and a deep supervision strategy accelerate model convergence.
• The model was trained and validated on TN3K, DDTI, and a large-scale clinical dataset.

Main Results:

• RTS-Net achieved superior performance in both in-distribution and cross-dataset evaluations.
• On TN3K, it reached 81.66% F1-score and 71.87% IoU; on DDTI, 71.10% F1-score and 60.09% IoU.
• Outperformed state-of-the-art methods including UNet, DeepLabv3+, and TransUNet.

Conclusions:

• The dual-path attention and graph convolution modules effectively improve feature representation and boundary delineation.
• RTS-Net demonstrates strong generalization, particularly for small nodules and blurred edges.
• Future work may involve integrating foundation models for enhanced robustness against artifacts.