Layer Segmentation Technology Of Optical Coherence Tomography Image Based On Deep Learning

Optical coherence tomography(OCT)is widely used in retinal clinical examination.It can obtain high-resolution retinal cross-sectional scanning images.Retinal and choroidal layer thickness is the key index for the study of ophthalmic diseases and other systemic diseases,the accurate measurement of which is very important for clinical analysis and early diagnosis of diseases.At present,a large number of deep learning algorithms have been proposed to segment the OCT retinal image,but most of these methods ignore the correlation between image features.They lack the use of global context semantic information and are limited to full convolution processing,which can not build a long-term correlation between the image features,and the segmentation ability of the model is limited.It’s necessary to develop a algorithm to segment the retinal and choroidal layer more accuratly.In addition,OCT images taken by different devices have significant quality differences,and their fields are often different.When the supervised learning segmentation model trained on the source domain OCT image is used to segment the target domain OCT image,the segmentation generalization ability of the model is greatly decreased.Supervised learning method training is only applicable for labeled images,but it costs a lot of labor and time on annotating OCT images requires.Therefore,it’s necessary to develop a domain adaptation method based on a high-precision segmentation algorithm.It can use labeled source domain images to achieve more accurate segmentation results for unlabeled target domain images,and expand the applicable scope of the segmentation algorithm.The domain adaptation method can reduce the workload of doctors and provide clinical reference for doctors,it has important practical significance.Aiming at the accurate layer segmentation of OCT image,a Multi-scale feature Fusion layer segmentation Network based on Transformers(MFTrans-Net)is proposed to segment OCT image with supervised learning.Combined with deep residual network,a multi-scale feature fusion module was proposed,the feature maps of different scales and grades are weighted and fused.The fusion features not only contain more location and detail information of low-level features,but also have stronger semantic information of high-level features.Then,the self-attention mechanism is used to construct the long-range correlation on the fusion features and capture the internal correlation between features at different scales.In the decoding stage,the encoded fusion features are restored to the original scale and injected into the decoder.This process above The global context information and the correlation between semantics of different scale features are fully considered in the process above.The segmentation experimental results show that the performance of MFTrans-Net is better than all competed methods.And the method proposed can achieve more accurate layer segmentation of OCT images.Aiming at the domain adaptation problem of OCT image layer segmentation,an unsupervised Selective-Guided Adversarial Adaptation(SGAA)is proposed in the thesis to segment the unlabeled target domain OCT images.The algorithm constructs a dual encoder structure to encode the source domain and target domain images in the same feature space.This structure can ensure that the coding paths of different domain images are independent of each other,and decouple the coding parameters to avoid the parameter entanglement caused by the full sharing of the source and target domain network.Then a parameter selective guidance strategy is proposed.Combined with the standards of parameter weight and gradient,the parameters,suitable for target domain OCT image coding in the source domain encoder,are selected and transferred to the target domain encoder.In the output space of the network,the feature distance between the source and target domain is shortened by using adversarial learning,so that the target domain encoder can update continuously the adaptive parameters to achieve continuous progress.The experimental results of domain adaptation show that SGAA can overcome the domain shift significantly and the performance of SGAA is better than other state-of-the-art domain adaptation methods.Finally,MFTrans-Net segmentation algorithm combined with SGAA adaptation method is used in this thesis to segment a batch of OCT images which contain White Matter Hyperintensities(WMH),Parkinson’s Disease(PD)and Healthy Control(HC),and extract the thickness of each outer retinal layer to analyze the disease correlation.The results of statistical analysis show that the thickness changes of some outer retinal sublayers are correlated with WMH and PD,which can provide clinical reference for doctors,it also provides the possibility of diagnosing WMH and PD through OCT images.