Deep-Learning Based,Recognition Of Retinopathy In Optical Coherence Tomography Image

By the end of 2017,the number of diabetic patients in the world has reached 425 million.The risk of diabetic retinopathy become higher with the prolongation of the diabetes duration.And this disease has become the main cause of visual impairment and even blindness in diabetic patients.For example,diabetic macular edema is a common symptom that threatens the vision of diabetic patients.Besides,macular degeneration is another common cause of blindness.The early diagnosis and treatment of these diseases is the key to prevent the deterioration of vision.However,such a large number of screening and diagnostic work will inevitably require a considerable number of professional ophthalmologists,while lack of doctors in China's first-tier cities and even in developed countries.This problem is more severe in the rural areas and other underdeveloped areas.In view of the above problems,the main contribution of this thesis is to apply advanced computer vision technology to medical image recognition task,and to construct an intelligent recognition system which can be used for medical image automatic screening and diagnosis.It can be used for on-line diagnosis,which is convenient for doctors and patients.In the retinopathy recognition task,we mainly use two open source reti nal OCT image data sets,their data labels are manually tagged by professional ophthalmologists.Among them,2014_BOE_Srinivasan includes 3231 OCT images,and the classification tasks are AMD,DME and NORMAL.OCT2017 includes 84,484 OCT images and the classification tasks are CNV,DME,DRUSEN and NORMAL.Based on these data sets,this thesis trained four optimized convolutional neural networks then compared and analyzed their recognition results.Among them,the Xception model with 126 layers achieved the best recognition accuracy.In the 2014_BOE_Srinivasan,the recognition rate for AMD,DME and NORMAL were all 100%.In the OCT2017,the accuracy,recall,sensitivity,and specificity of the Xception model were 97.5%,97.5%,98.39%,and 98.20%,respectively.The recognition rates of CNV,DME,DRUSEN and NORMAL are 98%,97.2%,95.6% and 99.2%,respectively,which are superior to the best recognition results published in the existing literature.Based on the high-accuracy classification results,this thesis combined with the open source U-net model to segment the image of the DME,in order to improve the interpretability of the model diagnosis results.Through comparison,we find that the recognition accuracy of our model has reached the level of professional ophthalmologists.In addition,in order to verify the reliability of the depth model in the medical image recognition task,this thesis uses the transfer learning method to migrate the Xception model to the pediatric pneumonia chest X-ray image recognition task.In the recognition of two types of tasks,the model finally achieved an accuracy of 96.31% and a recall rate of 95.94%.In subdividing the three types of identification tasks,the model achieved an accuracy of 88.78% and a recall rate of 86.87%.