Researches On Fundus Image Quality Assessment And Optimization To Assist Medical Diagnosis

Fundus images captured by fundus cameras,including retinal optic disc,macula and retinal blood vessel structures,are one of the most essential medical images and are widely used in assisting medical diagnosis.On the one hand,they can assist ophthalmologists for the diagnosis of fundus diseases such as senile macular disease,glaucoma and diabetic retinopathy.On the other hand,since retinal blood vessels are the only vascular structures that could be observed non-invasively in human blood circulation system,retinal vessels can also be applied to assist doctors in diagnosing some systemic cardiovascular diseases.Thus,in order to make fundus images to assist medical diagnosis better,this thesis focuses on the researches of fundus image quality assessment and blood vessel optimization concerning the following contents:(1)An automated quality assessment method of fundus image via brightness,image naturalness and structure analysis is proposed,combining image brightness,naturalness and structure indexes to classify fundus images into good or poor quality categories automatically,images with poor quality are recommended to re-capture.Firstly,1000 representative fundus images are selected to build the assessment database,and subjective scores are rated based on the density,naturalness and the position of optic disc.Then,a blind objective quality evaluation model is established: Pixel value statistics are used to evaluate image brightness;Multivariate Gaussian model is trained by multiple features to predict image naturalness and clearness.And the position of fundus optic disc is obtained to evaluate the image structure.Finally,overall quality of fundus image can be obtained by multiple quality regression methods.(2)A blood vessel segmentation method by cross-modality dictionary learning is proposed.Firstly,considering some fundus images are of low contrast and inhomogeneous brightness,we conduct the optimization preprocessing: we balance image brightness by gamma correction and obtain the vascular enhanced image by multi-scale Frangi filtering.Then,KSVD algorithm is applied to train enhanced image blocks and the corresponding label blocks to obtain the description dictionary and segmentation dictionary.And the cross-modality relationship between enhanced images and retinal blood vessels can be built to segment retinal blood vessels.Finally,segmentation results are acquired by post-processing operations like noise removal and hole-filling,and this paper evaluate the segmentation performance by eight metrics such as accuracy,specificity and sensitivity in three available databases HRF,DRIVE and STARE.(3)A discriminative retinal blood vessel segmentation method based on the fusion of multiple features is proposed.Firstly,six kinds of enhancement algorithms are employed to obtain six different vascular enhanced maps that will be divided into uniform image blocks.Each image block is labeled with a thick or thin vessel classification label to improve detection accuracy of tiny vessels.Then,LC-KSVD algorithm framework is used to train these enhanced vascular blocks,vessel classification labels and the corresponding ground-truth label blocks,so that the transformation matrix and reconstruction dictionary can be acquired to reconstruct and segment blood vessels.Finally,segmentation results are obtained by post-processing like noise removal and hole-filling.Segmentation performance evaluated in HRF,DRIVE and STARE shows this method can segment small vascular structures better and have greatly improved the sensitivity.