Research On Multi-disease Classification Method Of Fundus Image Based On Deep Residual Network

Fundus images can reflect some lesion features,which have become an important basis for ophthalmologists to judge the fundus disease of patients.During the collection of fundus images,involuntary eye movement,uneven illumination,and camera shake can all lead to poor image quality and difficulty in distinguishing lesion features.This paper focuses on helping doctors accurately screen and diagnose fundus diseases,as well as achieving precise processing of patients’ fundus images and realizing automatic diagnosis of multiple eye diseases.The research mainly includes:(1)An image enhancement algorithm based on the HSL color space,named HSLexNet,was proposed to address the issues of uneven illumination,low contrast,and blur in retinal images.The algorithm involves converting the RGB image to the HSL color space and utilizing the Retinex-Net algorithm on the luminance component to eliminate the effects of illumination variations and enhance the brightness,contrast,and details of the image.A reconstruction loss function is defined based on the Retinex theory,which reconstructs the luminance component(L channel)of low-illumination fundus images to match that of normal-illumination fundus images.Additionally,a cross-entropy loss function is utilized to optimize the reconstructed images.Experimental validation is conducted on the Diaretdb and LOL datasets,and HSLex-Net demonstrates its advantages with respect to grayscale mean,mean value,and average gradient,achieving values of91.625,82.528,and 25.426,respectively.(2)The classification performance of deep residual networks for retinal images is enhanced by employing a novel feature extraction method based on uniqueness and compactness.Due to significant distinctions between abnormal lesion features and surrounding normal tissues,the uniqueness and compactness characteristics of lesion features are utilized to accurately extract the contours of lesions.Subsequently,the extracted lesion contours are incorporated as constraints into the classification network,together with retinal images,during the training of the classification model,thus improving the classification performance.(3)A method for multi-disease classification of retinal images based on an improved Res Net model is proposed.The enhanced images are used as the basis,and the lesion contour maps are integrated as inputs to the Res Net network.A multi-branch structure is adopted for disease classification.Attention mechanisms are applied within the residual blocks,after the global pooling layer,and between different residual blocks.The model fusion is performed using an average fusion method to improve classification accuracy.Experimental results show that the accuracy of the Pool＿Attention is 0.751,with an AUC of 0.905,an F1-score of 0.904,and a Jaccard index of 0.685.The fusion of the M＿Res Net50 model and the Pool＿Attention model improves the classification accuracy,with a fusion accuracy of 0.779,an AUC of 0.913,an F1-score of 0.905,and a Jaccard index of 0.702.The proposed Retinex-Net enhancement algorithm based on the HSL color space in this paper outperforms other comparative methods in terms of clarity improvement,brightness enhancement,and noise suppression.The extracted lesion features using the joint saliency feature extraction method assist the classification network in identifying lesion characteristics.The improved Res Net-based multi-disease classification method for retinal images exhibits significant advantages in terms of classification accuracy and robustness.It provides valuable assistance to clinical doctors in diagnosis and has a certain auxiliary role.