Biological Vision Inspired Retinal OCT Image Segmentation

Quantitative analysis of the thickness of retinal layers is of great significance in ophthalmology for estimating pathological changes and diagnosing retinal diseases.The thickness of retinal layers can be acquired by retinal optical coherence tomography(OCT)image segmentation.Therefore,developing the technique of retinal OCT image segmentation is critical important.Thereare some difficult for distinguishing the retinal layer structure,like speckle noise caused by light scattering in OCT imaging process,which can degrade the image quality,and low contrast between adjacent layers due to small differences in the reflective coefficient and tissue uneven.In the past,the identification of retinal layer boundaries was usually performed manually by ophthalmologist.Howevear,manual segmentation is both time-consuming and suffers from inter-and intra-rater variability.With the improvement OCT technology,it is necessary to develop computer-assisted retinal layer segmentation method.Human visual system is the earliest known and the most powerful and sophisticated visual system,which has excellent information processing capabilities that enable quick and efficient extraction of useful information from a large number of complex scenarios.In particular,the visual system can correctly extract the object contour from cluttered background,underlying this capability is contour integration mechanism,which can appropriately link local edge elements into global contours.According to the accounts of contour integration process developed from Gestalt"good continuity" rules,the visual contour integration can be summarized using three processing stages.1)Contour elements extraction,just like edge detection,2)computing the "binding strengths" between local contour elements,3)using the local binding strengths to integrate the local elements into global contours.Inspired by the biological vision,a retinal OCT segmentation method based on visual contour integration is proposed to solve the problem of false edge information caused by noise and edge broken due to tissue uneven in OCT image edge detection task.Combined with the features of the boundaries of retinal layers,using the parameter model that simulates the process of human visual contour integration to calculate the linking probability between two edge fragments.Then,whether the two separate edge elements are derived from one same contour is judged from the linking probability.An innovative edge linking algorithm was developed.Combined with an accurate edge detection method for OCT image,the proposed edge linking algorithm can extract these edge elements belonging to the same physiological layer boundary from one retinal layer boundary from a cluttered edge image and integrate them into a complete contour,and achieve the retinal OCT image segmentation.The segmentation of retinal OCT images that connected with human visual characteristics,can make the segmentation results more in line with the judgment of human eyes,and improve the accuracy of segmentation.In experiment part,154 normal retinal SD-OCT images have been segmented using the proposed method.Results show that ten retinal layer boundaries can be extracted completely to achieve the thickness of nine retinal layers with success rate of greater than 90%by the contour integration-based method.Quantitative analysis was carried out by comparing with the boundaries labeled by two ophthalmologists,Boundary positioning error and layer thickness differences both are less than one pixel.However,the DSC values show that the coincidence rate between the segmentation by proposed method and the manual segmentation is only about 50%,even though the manual segmentation comparison,it hard to achieve complete coincidence in the segmentation results by different people because of the inter-rater variability.To exclude the effect of subjectivity on the assessment results,this paper uses the known boundaries as the ground truth,and uses them to synthesize retinal OCT-like images,then apply the proposed segmentation method to delineate the boundaries of retinal OCT-like images.The quantitative results show that the proposed method has better accuracy,both boundaries positioning error and layers thickness differences are less than 0.5 pixels,DSC values show that the coincidence rate is greater than 85%.In summary,the retinal OCT images segmentation method proposed in this paper has higher accuracy.It takes advantages of the characteristics of the information processing in human visual system and develops a computer-implemented visual bionic model.Comparing with these traditional edge linking algorithms using neighborhood interpolation,the edge linking algorithm based on the contour integration mechanism of human visual system can integrate these edge elements with far physical distance but has good continuity in the global contour,and can eliminate false edges effectively.