Classification Model of Diabetic Retinopathy Based on a Lightweight Feature-Enhanced Residual Swin Transformer
2
조회
0
다운로드
0
인용
0
참고문헌
키워드
초록
The automated detection of diabetic retinopathy (DR) relies heavily on retinal image analysis. While artificial intelligence models have shown promise in DR management, they often face challenges such as high computational complexity, reduced accuracy due to class imbalance and small inter-class gaps, and increased processing times. Addressing these limitations, this study introduces the lightweight feature-enhanced residual Swin (LFRS) Transformer, a model that maintains high accuracy despite significantly lowered computational demands. Our approach begins by converting color fundus images to grayscale, followed by local feature extraction using a depthwise separable convolution module. These features are subsequently subjected to processing by a lightweight Swin Transformer enhanced with residual connections, improving both global feature extraction and computational efficiency. Evaluated on the DR classification dataset released by APTOS 2019, the LFRS Transformer achieves an accuracy of 0.928, a recall of 0.965, and a weighted kappa score of 0.957. Compared to the baseline Swin Transformer, our model reduces computational load by 22.2 GFLOPs and decreases model parameters by 7.2M, demonstrating a substantial improvement in efficiency. These results underscore the LFRS Transformer as a highly efficient and reliable DR screening tool, positioning it as well-suited for large-scale clinical screening programs.
Abstract
The automated detection of diabetic retinopathy (DR) relies heavily on retinal image analysis. While artificial intelligence models have shown promise in DR management, they often face challenges such as high computational complexity, reduced accuracy due to class imbalance and small inter-class gaps, and increased processing times. Addressing these limitations, this study introduces the lightweight feature-enhanced residual Swin (LFRS) Transformer, a model that maintains high accuracy despite significantly lowered computational demands. Our approach begins by converting color fundus images to grayscale, followed by local feature extraction using a depthwise separable convolution module. These features are subsequently subjected to processing by a lightweight Swin Transformer enhanced with residual connections, improving both global feature extraction and computational efficiency. Evaluated on the DR classification dataset released by APTOS 2019, the LFRS Transformer achieves an accuracy of 0.928, a recall of 0.965, and a weighted kappa score of 0.957. Compared to the baseline Swin Transformer, our model reduces computational load by 22.2 GFLOPs and decreases model parameters by 7.2M, demonstrating a substantial improvement in efficiency. These results underscore the LFRS Transformer as a highly efficient and reliable DR screening tool, positioning it as well-suited for large-scale clinical screening programs.