Research On Vessel Quantization And Image Processing Of AS-OCTA Images Based On Deep Learning

Medical imaging is an important auxiliary means for clinical diagnosis.Optical Coherence Tomography Angiography(OCTA)is a new non-invasive imaging technology,which can clearly show the vascular morphology of the eye.The vascular information in the image provides an important basis for disease detection,diagnosis,and treatment.It has become a research highlight to process and analyze medical image data using deep learning methods to achieve intelligent auxiliary diagnosis and treatment.However,due to the complex structure of the eye and the periodic horizontal stripe noise generated in the image caused by the involuntary eye tremor,accurate segmentation and quantification of blood vessels in ophthalmic images based on deep learning technology is still a very challenging task.In this thesis,the following research is carried out around the key problems in the extraction and quantification of blood vessels in the Anterior Segment(AS)OCTA images:(1)In order to solve the noise problem of AS-OCTA images,this thesis concentrates on the OCTA image enhancement algorithm based on deep learning.Firstly,an original image is composed of noise information and clean image information.The reconstruction term is constrained by the original image,noise,and clean image.Secondly,the constraint of the stripe term is constructed by using the low-rank feature of the stripe noise distribution in the image.Then,according to the anisotropy of the vascular structure,an anisotropic total variation restriction term is constructed to ensure the integrity of the vessel structure after noise removal.Finally,the stripe loss function is constructed by the above three terms.The Stripe Removal Network(SR-Net)is proposed based on the destripe loss function.The de-stripe model is trained by SR-Net.Experiments were performed on the synthetic retinal OCTA dataset and the real AS-OCTA dataset,and the model achieved good image enhancement results.The PSNR and SSIM have been significantly improved in the test results of the synthetic retinal OCTA dataset.(2)In order to extract the vessel structure from AS-OCTA accurately,the Cascaded Noise Suppression Segmentation Network(CNSS-Net)is proposed in this thesis.Due to the serious noise problem of the AS-OCTA dataset,CNSS-Net incorporates Subspace Attention(SSA)as a noise suppression module and combines it with several micronetworks.The SSA module can suppress the influence of the noise on segmentation results,and the cascade network architecture can improve the continuity of vascular segmentation results.Finally,CNSS-Net achieves accurate segmentation results for ASOCTA images with more noise,which significantly improved in the Sen and Dice.(3)This thesis research on vascular quantification methods aims to provide reliable evidence for disease screening and clinical auxiliary diagnosis.Based on the proposed image enhancement(SR-Net)and image segmentation(CNSS-Net)in this thesis,we research that the image segmentation results will be improved after the image enhancement.Experiment results show that the noise images after SR-Net will get more accurate vessel segmentation results.Moreover,the proposed CNSS-Net can get better vascular segmentation results in the noisy AS-OCTA images.It provided the foundation for the accurate quantification of biological information.The vascular information was extracted after SR-Net enhancement and CNSS-Net segmentation.The vascular density(VD)index used to measure the vascular intensity information is obtained using the area method calculated by pixel counting,and the fractal dimension(FD)index used to measure the vascular structure information is obtained using the box counting method.The experimental results show that the synthesized retina datasets ROSE-1 and PUTH reach 29.025,0.945 and 29.117,0.949 on PSNR and SSIM respectively after SRNet image enhancement.After the CNSS-Net blood vessel segmentation of AS-OCTA,Dice and Sen reached 65.6% and 74.4% respectively.After the image enhancement and image segmentation methods proposed in this thesis,accurate results of vessel segmentation have been obtained,and based on the results of vessel segmentation,the vessel information of dry eye patients has been successfully extracted and quantified,which provides a basis for clinical diagnosis of doctors.