Related Experiment Video
Updated: Jul 26, 2025
A Swin Transformer-Based Model for Thyroid Nodule Detection in Ultrasound Images
Published on: April 21, 2023
FBN: Weakly Supervised Thyroid Nodule Segmentation Optimized by Online Foreground and Background.
Ruiguo Yu1, Shaoqi Yan1, Jie Gao1
1College of Intelligence and Computing, Tianjin University, Tianjin, China; Tianjin Key Laboratory of Advanced Networking, Tianjin University, Tianjin, China; Tianjin Key Laboratory of Cognitive Computing and Application, Tianjin University, Tianjin, China.
This study introduces a new method for segmenting thyroid nodules in ultrasound images using only classification data, significantly improving accuracy. The approach enhances segmentation performance by leveraging image information, bridging the gap between weakly and fully supervised methods.
Area of Science:
- Medical imaging
- Artificial intelligence
- Computer vision
Background:
- Semantic segmentation of thyroid nodules is crucial for diagnosis.
- Pixel-level annotation of ultrasound images is labor-intensive.
- Weakly supervised semantic segmentation (WSSS) methods often struggle with complete object highlighting.
Purpose of the Study:
- To train a semantic segmentation model for thyroid nodule ultrasound images using classification data.
- To reduce the need for extensive pixel-level labeled datasets.
- To improve segmentation performance by mining image information and bridging the WSSS-FSSS gap.
Main Methods:
- Proposed a novel foreground and background pair (FB-Pair) representation method.
- Utilized Class Activation Maps (CAM) and revised them using FB-Pair.
- Designed a self-supervised learning pretext task based on FB-Pair for accurate object distinction.
Main Results:
- Achieved a 5.7% improvement in mean intersection-over-union (mIoU) segmentation performance.
- Reduced the performance difference between benign and malignant nodule segmentation to 2.9%.
- Demonstrated superior performance compared to existing methods on the thyroid nodule ultrasound image (TUI) dataset.
Conclusions:
- Successfully trained a high-performing segmentation model using only classification data for thyroid nodule ultrasound images.
- Confirmed that CAM can effectively utilize image information for accurate target region highlighting, thereby enhancing segmentation.
- The proposed method offers a viable solution for efficient and accurate thyroid nodule segmentation.
