DPF-EHDNet: a differential-path and structurally enhanced network for thyroid ultrasound segmentation.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Ultrasound Technology

Background:

• Accurate segmentation of thyroid nodules in ultrasound is difficult due to image noise and variable lesion characteristics.
• Reliable segmentation is crucial for clinical assessment and computer-assisted diagnosis (CADx) systems.

Purpose of the Study:

• To develop a robust segmentation framework, DPF-EHDNet, specifically designed for challenging ultrasound conditions.
• To improve the accuracy and consistency of thyroid nodule segmentation in the presence of speckle noise and low-contrast margins.

Main Methods:

• Proposed DPF-EHDNet, a novel deep learning framework incorporating differential-path feature enhancement, edge-aware multi-scale context encoding, and confidence-guided shallow feature fusion.
• Trained the model using a fixed-iteration protocol for fair comparison and evaluated performance averaged over three random seeds.
• Utilized a combined dataset from three public benchmarks: DDTI, TN-SCUI2020, and TN3K.

Main Results:

• DPF-EHDNet achieved high segmentation performance, with mean Intersection over Union (mIoU) of 93.09%, Dice coefficient of 93.35%, precision of 93.98%, and recall of 92.73%.
• The proposed model consistently outperformed existing baseline segmentation methods on the challenging thyroid ultrasound dataset.
• Demonstrated robustness in handling speckle degradation and low-contrast conditions inherent in ultrasound imaging.

Conclusions:

• DPF-EHDNet offers a robust and structurally consistent solution for thyroid ultrasound segmentation, significantly improving accuracy under difficult imaging conditions.
• The framework shows potential for enhancing clinical workflows and advancing computer-assisted ultrasound analysis for thyroid nodule detection and characterization.