A Multi-Perceptual Learning Network for Retina OCT Image Denoising and Classification

Swept Source Optical Coherence Tomography (OCT), a non-invasive cross-sectional imaging technique, has been widely used in diagnosing and treating various vision-related diseases. However, OCT images often suffer from heavy noise issues, due to the limitations of imaging devices, making analysis and disease classification a great challenge. This paper proposes a Multi-Perceptual Learning Network (MPLN) for retina OCT image denoising and classification. We adopt a triplet cross-fusion GAN approach and use three unpaired OCT images to conduct perceptual learning. In addition, we integrate the Frequency Distribution Loss into GAN to preserve both the structural integrity and perceptual quality of the denoised OCT images, enabling better classification. The method can significantly reduce the noise of highly noisy images. Our proposed method is evaluated on the VIP Cup 2024 dataset in terms of the CNR, MSR, and TP scores. Our model achieves a CNR score of 6.351, and an MSR score of 11.573, which outperforms many existing methods on OCT images. In classification, our MPLN improves accuracy by more than one percent. These results demonstrate that our model can significantly enhance image quality and improve classification accuracy, highlighting its potential for clinical applications.