Related Experiment Video
Updated: Jan 10, 2026
A Swin Transformer-Based Model for Thyroid Nodule Detection in Ultrasound Images
Published on: April 21, 2023
Dual-Scale Swin Transformer via Feature Alignment and Adversarial Discrimination for Retinopathy of Prematurity
Insights
A new dual-scale Swin Transformer (DS-Swin-T) network improves retinopathy of prematurity (ROP) diagnosis by reducing image style variations. This AI model enhances classification accuracy for premature infants, aiding early detection and treatment.
Area of Science:
- Ophthalmology
- Medical Imaging
- Artificial Intelligence
Background:
- Retinopathy of prematurity (ROP) is a significant cause of childhood blindness in premature infants.
- Effective management relies on timely diagnosis and treatment, which can be hindered by variations in medical image styles.
- Current AI models may face performance degradation when applied to datasets with different image styles.
Purpose of the Study:
- To develop a robust deep learning model for ROP classification that is invariant to image style variations.
- To mitigate the performance drop typically observed when transferring models trained on one image style to another.
- To enhance the accuracy and reliability of automated ROP detection systems.
Main Methods:
- Proposed a dual-scale Swin Transformer (DS-Swin-T) network incorporating image synthesis (IS), feature alignment, and adversarial learning.
- The IS module generates intermediate style images to reduce style discrepancies.
- Feature alignment and adversarial learning techniques were employed to extract style-invariant features for consistent classification.
Main Results:
- The DS-Swin-T network achieved 97.91% average accuracy on the source style dataset.
- When transferred to different target style datasets, the model maintained high performance, reaching a maximum average accuracy of 93.66%.
- The proposed method effectively addressed performance degradation caused by image style differences.
Conclusions:
- The DS-Swin-T network demonstrates significant effectiveness in classifying retinopathy of prematurity across varying image styles.
- This approach offers a promising solution for improving the generalizability and reliability of AI-based diagnostic tools in medical imaging.
- The findings highlight the importance of addressing style variations for robust clinical application of deep learning models.
Abstract:
Retinopathy of prematurity (ROP) is a retinal vascular disease that primarily affects premature infants with low birth weight. It is a leading cause of childhood blindness worldwide, but it can often be effectively managed with appropriate and timely diagnosis and treatment. To address the impact of image style on model classification performance, this paper proposes a dual-scale Swin Transformer (DS-Swin-T) network for ROP. The network comprises three components: image synthesis (IS), feature alignment, and advanced adversarial learning. The IS module generates synthesis style images as an intermediate latent space between source and target styles, reducing style difference. The DS-Swin-T serves as the primary framework for image feature extraction. Detail and style encoders extract features in the shallow feature space, with detail and style losses aligning these features to ensure consistency across styles. To extract rich style-invariant features and ensure consistent classification within the same category, adversarial learning is applied in the advanced feature space. Finally, feature fusion units process dual-scale classification representations. Our method achieves an average accuracy of 97.91% on the source style dataset. When transferred to other target style datasets, our method effectively mitigates the performance degradation caused by style difference, reaching a maximum average accuracy of 93.66%. Extensive experiments demonstrate the effectiveness of our method.
