Study Of Ophthalmic Image Classfication,Segmentation And Regression Based On Convolutional Neural Networks

With the development of modern society,more and more people are suffering from eye diseases.Though the number of patients with eye diseases is increasing year by year,the rate of growth of ophthalmologists is far behind the rate of growth of patients.As a result,developing a compter adied diagnosis system for eye diseases is urgently needed,which not only can alleviate the burden on the clinicians by providing objective opinion with valuable insights,but also can offer early detection and easy access for patients.In a computer adied diagnosis system for eye diseases,image classification,segmentation and regression play an important role.In this thesis,we do researches for these three issues based on convolutional neural networks(CNNs).For classification,we propose a surrogate assisted algorithm,which can augment the annotated data effectively and classify the retinal optical coherence tomography(OCT)images automatically.Concretely,image denoising is first performed to reduce the noise.Thresholding and morphological dilation are applied to extract the masks.The denoised images and the masks are then employed to generate a lot of surrogate images,which are used to train a CNN model.Finally,the prediction for a test image is determined by the average of the outputs from the trained CNN model on the surrogate images.The proposed method has been evaluated on different databases and achieves very competitive results,which demonstrates that the proposed method is a very promising tool for classifying the retinal OCT images automatically.For segmentation,we propose a novel algorithm by integrating the advantages of CNNs and parametric active contour models.Concretely,gradient vertor flow is first employed to train a CNN model.Once training done,the trained CNN is applied to derive an external force from an original image.The derived external force is then used to initialize the contours,and integrated into the active contour models for curve evolution.To demonstrate the effectiveness and versatility of the proposed method,we apply it on different segmentation tasks,including segmentation of optic disk in fundus images,fluid in retinal OCT images and fetal head in ultrasound images.The results show that the proposed method is very promising since it achieves competitive performance for different tasks compared to the state-of-the-art algorithms.For regression analysis,we employ CNNs to estimate the area of choroidal neovascularization(CNV)in OCT Images directly.Besides,we also explore how varying the input types of CNNs affects the performance.Some interesting findings are gained via this study.For example,the performance of the direct method is very competitive with(or even better than)the segmentation based methods,especially,for those cases that the CNV is diffictult to be segmented.In addition,we also find that different input types have different affection on the performance.Such as CNV enhancement is useful for improving the performance.Histogram counts would lead to the performance deteriorated significantly,while histogram of oriented gridients would not.These findings should be considered for improving the performance of CNNs in the future work.In summary,the study in this thesis can facilitate the development of computer adied diagnosis systems for eye diseases.Concretely,the proposed surrogate assisted algorithm can augment the annotated data effectively.The proposed segmentation algorithm can improve the accuracy of segmentation via integrating the advantages of CNNs and parametric active contour models.The study of employing direct methods for CNV area estimation can facilitate the application of direct methods on CNV area estimation.