Study And Application Of Segmentation Methods For Fundus Images

Image segmentation technology has always been a hot and difficult problem in medical image processing and analysis. Segmentation of early diabetic retinopathy(DR) lesions in fundus images belongs to weak signal extraction. The features of fundus image are complex and the ones of DR lesions are diverse etc., so that it is more challenging for segmentation of early DR lesions in fundus images. On the basis of deep research on the method of fundus imaging, features of anatomical structures in fundus and pathological knowledge about related lesions, meanwhile combined with clinical actual demand, did research on medical image segmentation technology which was applied to automated screening of DR by means of theory, algorithm and implementation. The main research contents are listed as follows:(1) Spectral signature of major anatomical structures in fundus was studied, so that the right channel could be selected for different segmentation objects which provided a good basis for subsequent segmentation.(2) Segmentation methods of hard exudates which is clinical signs of early DR in fundus images were studied. Firstly, optic disc as a false positive relative to hard exudates in fundus images was segmented by the method of locating and extracting it according to the direction of main blood vessel network which was on the basis that main blood vessel converged in optic disc. Then, two segmentation methods of hard exudates, one based on mathematical morphology and the other based threshold segmentation and pattern classifiers, were proposed and investigated.(3) Segmentation methods of bright lesions including hard exudates and cotton wool spots which are main symptoms of early DR in fundus images were studied. To obtain candidate regions of bright lesions, improved FCM(IFCM) algorithm was established and studied. In this improved FCM algorithm, clustering centers of FCM were initialized by the results of K-means, as a result that iterations of FCM algorithm would be significantly reduced and clustering speed of FCM was significantly improved. In addition, median filter was added to the criterion function of FCM which directly put the noise suppression into the updating formulas of fuzzy partition matrix U and clustering centers V, so as to efficiently raise accuracy of U and V, Thus IFCM algorithm could effectively overcome the shortcoming of sensitivity to noise. Then, an automatically detecting approach based on IFCM and SVM which was used to detect and classify bright lesions of DR from color fundus images was proposed. In this approach, A two-level SVM classification structure whose advantage was that unlike other schemes, users can access the results of each classifier for clinical consideration was applied to classify the candidate regions obtained by IFCM algorithm. As a result, this approach met the clinical demands better.(4) Segmentation methods of hemorrhages which is another main symptom of early DR in fundus images were studied. An algorithm of locally adaptive region growing based on multi-template matching was established and studied on the basis that theory as well as advantages and disadvantages of template matching and region growing were studied. Normalized cross-correlation(NCC) template matching used in the new algorithm was improved calculation efficiency by the introduction of convolution operation. Locally adaptive region growing was that regions of seeds growing were limited in a local scope by each target, and adaptive parameters were added into growing rules of region growing. This novel algorithm named locally adaptive region growing based on multi-template matching is suitable for extracting several objectives of the same kind in images whose shape was certain or not certain. The proposed algorithm was used to detect hemorrhages in fundus images, results suggest that the approach could fast and effectively detect hemorrhages in fundus images, and it is stable and of high precision and efficiency.(5) Segmentation methods of microaneurysms which is one kind of DR clinical signs were studied. microaneurysm is a reddish, circular pattern with a diameter less than 125 μm according to the medical definition. So it could be found by its diameter and isolated connected red pixels with a constant intensity value, and whose external boundary pixels all have a higher value in the green plane of a RGB fundus image. Combined morphological characteristics of microaneurysm in the green plane and image characteristics characterized by extended-minima transform of mathematical morphology, a novel, simple and efficient approach of automatic microaneurysms detection was established and studied on the basis that hard exudates segmentation had been accomplished before. Because it makes full use of morphological characteristics of microaneurysm, this approach is stable, reliable and of high precision as well as efficiency.An automated DR screening system based on non-dilated fundus images was developed on the research achievement above and tested on an image dataset collected in clinic. With an exam-based criterion, sensitivity of 96.46%, specificity of 96.07% are achieved. Results suggest that the performance of DR screening system meet the minimum standard of 80% sensitivity and 95% specificity of detection of sight-threatening DR which was recommended by British Diabetic Association guidelines.