[Thyroid nodule segmentation method integrating receiving weighted key-value architecture and spherical geometric...

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Context:

• Transformer models face challenges in ultrasound thyroid nodule segmentation due to high computational complexity.
• Traditional sampling methods can lose critical details in high-resolution ultrasound images, impacting nodule segmentation.
• Accurate segmentation of thyroid nodules is crucial for diagnosis and treatment planning.

Purpose:

• To develop a novel thyroid nodule segmentation method addressing computational complexity and detail loss in ultrasound images.
• To integrate the receiving weighted key-value (RWKV) architecture with spherical geometry feature (SGF) sampling for enhanced segmentation.
• To introduce a patch attention module (PAM) for optimizing feature map focus on high-resolution encoder features.

Summary:

• The proposed method utilizes RWKV architecture and SGF sampling with 2D offset prediction and pixel-level adjustment for precise segmentation.
• A patch attention module (PAM) refines decoder features via regional cross-attention, improving focus on encoder details.
• Experiments on TN3K and DDTI datasets show superior Dice Similarity Coefficients (DSC) of 87.24% and 80.79% respectively, with lower computational cost.

Impact:

• Achieves state-of-the-art performance in thyroid nodule segmentation with reduced computational complexity.
• Offers a more efficient and accurate solution for segmenting complex ultrasound thyroid nodule images.
• Potential to improve diagnostic accuracy and clinical workflow for thyroid nodule assessment.