TNSNet: Thyroid nodule segmentation in ultrasound imaging using soft shape supervision

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Endocrinology

Background:

• Thyroid nodules are common endocrine disorders.
• Accurate segmentation of nodules in ultrasound images is crucial for diagnosis but challenging due to image quality.
• Deep learning offers potential solutions for improving thyroid nodule segmentation.

Purpose of the Study:

• To develop and evaluate a deep learning algorithm for accurate thyroid nodule segmentation on ultrasound images.
• To improve the detection and segmentation of nodule boundaries using soft shape supervision.
• To enhance computer-aided diagnosis systems for thyroid nodules.

Main Methods:

• A dual-path convolution neural network (CNN) with DeepLabV3+ as the backbone was proposed.
• Soft shape supervision blocks were integrated to enable cross-path attention mechanisms.
• The network enhances shape features and incorporates them into the region path for improved segmentation.

Main Results:

• The network was trained and tested on 3786 ultrasound images.
• Achieved 95.81% accuracy and 85.33% Dice Similarity Coefficient (DSC).
• Demonstrated superior segmentation performance and accurate boundary learning compared to classical deep learning methods.

Conclusions:

• The proposed dual-path network accurately segments thyroid nodules in ultrasound images.
• This method serves as a valuable initial step for computer-aided diagnosis systems.
• The network shows significant potential for clinical application in nodule segmentation.