AFF-UNet-RWKV: A Lightweight Model for High-Quality Deblurring in Medical Imaging

Abstract

Medical image deblurring is essential for enhancing image quality and improving diagnostic accuracy. This paper introduces a lightweight deep learning model, AFF-UNet-RWKV, which combines AFF-UNet and the RWKV-lite spatial mixer for medical image deblurring tasks. By incorporating Attention Feature Fusion (AFF) and the RWKV-lite module, this model effectively integrates features from both the encoder and decoder while capturing long-range spatial dependencies, thereby enhancing the restoration of image details and structures during the deblurring process. Additionally, the model adapts its feature fusion strategy to maintain high recovery performance across various types of blur. To validate its effectiveness, this study conducted experiments using the PathMNIST subset from the MedMNIST dataset, generating blurred image pairs through Gaussian blur and linear motion blur for training. The results demonstrate that the AFF-UNet-RWKV model significantly outperforms traditional deblurring methods and other deep learning models in terms of image recovery quality. Notably, the approach shows substantial advantages over traditional algorithms and DeblurGAN, particularly in the Structural Similarity Index (SSIM). Ultimately, the model achieved a PSNR (Peak Signal-to-Noise Ratio) of 32.03 dB and an SSIM(Structural Similarity Index) of 0.898 on the test set, confirming its superiority and potential in medical image deblurring tasks. This research offers an efficient and lightweight solution for medical image deblurring, providing strong practical application value and new insights for future research and applications in related fields.