Research On The Prediction Of Posterior Capsular Opacification Based On Time Series Retro-illumination Images

Posterior capsular opacification is a common complication after cataract surgery,referred to as PCO.The appearance of PCO can directly affect the vision recovery of patients after cataract surgery.At present,Nd:YAG laser is an effective way to treat the PCO,but once the best treatment opportunity is missed,the posterior capsule can only be removed through the surgery.Therefore,the timely detection and treatment of PCO plays a key role in the recovery of vision in patients after cataract surgery.After cataract surgery,patients need to go to the hospital for review.The retro-illumination image is an effective medium for diagnosis and analysis of PCO.The images that are continuously reviewed can form time series images.By analyzing the time series retro-illumination images can discover the law of the development and changes of the PCO,so as to realize the prediction of the PCO.However,it is difficult for doctors to scientifically explore the laws of this.Therefore,this thesis proposes a PCO prediction method based on convolutional neural network(CNN)and convolutional long short-term memory network(ConvLSTM),which can generate the retro-illumination images of the eye at the next moment.And predict and assess its severity to determine whether patients need Nd:YAG laser treatment within the next six months.The main work of this thesis includes the following three aspects:(1)Research on automatic detection method of lens region.The clinical manifestation of PCO is opacity of the posterior capsule of the lens.Noise such as eyelashes and eyelids around the lens can greatly interfere with the feature extraction and prediction of PCO.Therefore,the pretreatment of retro-illumination image is studied first,a lightweight SSD network model based on deep learning is proposed,which can quickly and accurately detect the lens region.Its IOU positioning index is over 96%,and 57 images can be detected every second,which satisfies real-time requirements of the clinical situation.(2)Research on the prediction method of spatial-temporal series for automatically generating retro-illumination image at the next moment.In this thesis,an end-to-end model is established by combining convolutional neural network and ConvLSTM.By studying and analyzing the time series retro-illumination images of patients who have been reviewed several times after surgery,the high-level features of the images are extracted and the internal laws between time series features are mined.Finally,the retro-illumination image of the patient at the next moment is generated to help the doctor scientifically predict the area and position change state of PCO at the next moment.(3)Prediction of the severity of PCO based on generated images.Based on the residual convolutional neural network,a retro-illumination image grading model is built to automatically assess the severity of the generated image,and then predict whether the patient needs Nd:YAG laser treatment in the next six months,and the prediction accuracy is 92.35%.This method can help doctors plan treatments in advance to prevent patients from missing the best treatment opportunity.