Dense Swin Transformer for Classification of Thyroid Nodules

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Oncology

Background:

• Thyroid nodules are common and can be early signs of thyroid cancer.
• Ultrasound is a primary non-invasive method for diagnosing thyroid nodules.
• Distinguishing benign from malignant nodules via ultrasound is challenging due to subtle visual differences and low tissue contrast.

Purpose of the Study:

• To propose a novel deep learning method, the dense nodal Swin-Transformer (DST), for improved diagnosis of thyroid nodules.
• To enhance the accuracy and reliability of differentiating benign and malignant thyroid nodules using ultrasound imaging.

Main Methods:

• Image segmentation using a patch-based approach within the Swin-Transformer architecture.
• Construction of multi-scale feature maps across four stages to capture diverse feature information.
• Implementation of a dense connection mechanism within each stage block to leverage multi-layer features.

Main Results:

• The DST method achieved an accuracy of 87.27% on multi-center ultrasound data from 17 hospitals.
• Sensitivity reached 88.63%, and specificity was 85.16% in distinguishing thyroid nodule types.
• The algorithm demonstrated significant potential for assisting clinical diagnosis.

Conclusions:

• The proposed dense nodal Swin-Transformer (DST) method offers a promising advancement in AI-assisted thyroid nodule diagnosis.
• The DST method effectively utilizes multi-layer features for improved diagnostic performance in challenging ultrasound cases.
• This approach has the potential to support clinicians in making more accurate and timely decisions regarding thyroid nodules.