Research On The Segmentation Of Sugar Reticulum Lesions Based On Deep Learning

Diabetic retinopathy(DR),is a highly specific vascular complication of type I and type II diabetes,can lead to impaired vision and even blindness.In recent years,the spectral domain optical coherence tomography(SD-OCT)has a fundamental breakthrough in the field of imaging speed and resolution.It can clearly display 18 layers structure of retina and small lesions and collect hundreds of high-resolution tomographic images in a short time,which is of great significance to medical imaging and has become an important tool for ophthalmologists to diagnose fundus diseases.In this paper,the image processing and analysis are adopted to study the DR lesions in SD-OCT retinal images,including the following research contents:(1)An automatic segmentation method for hyper-reflective foci(HRF)based on image enhancement and improvement of 3D U-Net is proposed.At first,the raw images are denoised based on bilateral filter to reduce image noise.Some images show low HRFbackground contrast,which leads to the severe under-segmentation.Thus,we apply an enhancement algorithm over the whole dataset to enhance HRF in SD-OCT images.Then,the standard 3D convolutions at the last layer of 3D U-Net encoder path are modified into three different 2D dilated convolutions.With this structure,our network has more robust feature representations than other models by extracting 3D and 2D features.Finally,the denoised images and enhanced images are as the first channel and the second channel of the input to an improved 3D U-Net for training and testing.Experimental results demonstrate that the method proposed in this chapter can accurately segment the HRF in SD-OCT retinal images.(2)An unsupervised automatic segmentation method based on intensity anomaly probability for neurosensory retinal detachment(NRD)is proposed.Firstly,the mean value and standard deviation of the intensity of normal retinal images are obtained by layer segmentation,and the intensity distribution functions are obtained by Gaussian fitting.Then,with the help of layer segmentation and normal interlayer intensity distribution function,the probability maps are obtained by detecting pixel intensity abnormalities,and then preliminary segmentation results are obtained by thresholding.Instead of manual labeling,DUNet is trained using preliminary segmentation results.Finally,the softmax outputs of the network are optimized by using the fully connected conditional random fields.Experimental results show that the method proposed in this chapter can accurately segment NRD in SDOCT retinal images,which is more similar to ground truth visually.