Research On Pediatric Cataract Slit Lamp Images Diagnosis And Postoperative Effect Prediction Based On Deep Learning

Pediatric cataract is a typical pediatric ophthalmic disease and has been one of the most common diseases which affect the growth of children's vision.Pediatric cataract occurs at birth or at an early stage of childhood and can cause irreversible visual defects or blindness in serious cases if not treated in time.Pediatric cataract patients need to come back to the hospital for a regular check for a long time,because infant eyes are in the period of growth and pediatric cataract patients show wide variability in clinical.All above mentioned features not only result in sparse medical resources but also bring a heavy burden to the patients' families.With the development of machine learning,computer medical treatment has become an important way to assist doctors for medical diagnosis,which provides a new research direction to improve the current state of pediatric cataract.Slit-lamp images can intuitively reflect the morphology of lens opacity,and play a key role in the pediatric cataract diagnosis.Slit-lamp images supply rich data resources for computer aided diagnosis,which makes it possible for us to apply computer technology in pediatric cataract diagnosis.The research includes grading of pediatric cataract pre-operation oblique illumination images and prediction of post-operation retro-illumination time series images.In the grading task,locating and identifing lens in oblique illumination images by Canny edge detector and Hough transform is used firstly to eliminate noise.Then according to the grading standard which is formulated by ophthalmologists,three different CNN models are trained to grade pediatric cataract oblique illumination images from three perspectives which are lens opacity area,density and location.Besides,the area grading is taken as an example to optimize CNN model by testing the grading performance of CNNs with different parameters and network structures.Finally,the grading results of CNN and other methods that don't belong to deep learning are contrastively analyzed.In the prediction task,LSTM deep learning model is used to predict if the pediatric cataract patient needs to receive operation once more.And the prediction result of time series images by LSTM model is analyzed to explore dynamic mechanism of pediatric cataract recurrence.In the grading task,CNN classifier acquires good grading results in the three perspectives:lens opacity area,density and location,and accuracy rate reaches to 92.77%?89.55% and 76.63% respectively.And the grading results are better than the methods combining features extraction and SVM classifier.In the prediction task,prediction accuracy rate of LSTM reaches to 91.58%,which proves that pediatric cataract recurrence has its regularity and can be predicted.Previous work related to this thesis was published in the journal of Nature Biomedical Engineering,and was reported by IEEE Spectrum.As a whole,the results of CNN and LSTM models can serve as a reference for ophthalmologist to diagnose pediatric cataract.