A Swin Transformer-Based Model for Thyroid Nodule Detection in Ultrasound Images

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Thyroid cancer incidence is rising, making early thyroid nodule detection crucial.
• Convolutional Neural Networks (CNNs) show promise in thyroid ultrasound analysis but struggle with long-range contextual dependencies.
• Transformer networks excel at capturing long-range contextual information, essential for accurate nodule identification.

Purpose of the Study:

• To develop a novel thyroid nodule detection method combining Swin Transformer and Faster R-CNN.
• To address the limitations of CNNs in capturing long-range contextual information for thyroid ultrasound images.
• To improve the accuracy and sensitivity of thyroid nodule detection.

Main Methods:

• A hybrid approach integrating a Swin Transformer backbone with Faster R-CNN for thyroid nodule detection.
• Ultrasound images are projected into 1D embeddings and processed by a hierarchical Swin Transformer with shifted window self-attention.
• Feature fusion using a Feature Pyramid Network (FPN) followed by a detection head for bounding box and confidence prediction.

Main Results:

• The proposed method achieved a mean Average Precision (mAP) of 44.8%, outperforming CNN-based methods.
• Demonstrated superior sensitivity of 90.5% compared to existing approaches.
• Validated the effectiveness of context modeling for enhanced thyroid nodule detection.

Conclusions:

• The Swin Transformer and Faster R-CNN combination offers a powerful tool for thyroid nodule detection.
• Effective long-range context modeling is key to improving diagnostic accuracy in thyroid ultrasound.
• This method holds significant potential for early and accurate diagnosis of thyroid cancer.