Automatic Choroid Layer Segmentation in Optical Coherence Tomography Images Using Deep Learning
Roya Arian1 *, Tahereh Mahmoudi2 , Elias Khalili Pour3 , Hamid Riazi-Esfahani3 , Rahele Kafieh1
- Medical Image and Signal Processing Research Center, School of Advanced Technologies in Medicine, Isfahan University of Medical Sciences, Isfahan, Iran
- Department of Biomedical Systems & Medical Physics, Tehran University of Medical Sciences and Research Center for Science and Technology in Medicine, Tehran, Iran.
- Retina ward , Farabi Eye Hospital , Tehran University of Medical Sciences
Abstract: Choroid layer segmentation is essential task for analysis of a range of ocular diseases. Optovue Optical Coherence Tomography (OCT) has become a crucial tool in retinal imaging and provides noninvasive, high resolution cross-sectional view for retinal layers in vivo. OCT is preferred imaging technique that are essential for diagnosis and management of many retinal diseases are associated with the choroid. However, the boundaries between choroid and sclera in the OCT images is blurred and very difficult to be clearly detected. Manual segmentation of the large number of images of choroid is too time consuming, tedious with along human errors. Therefore, automated choroidal layer segmentation is pressing need and has received many attentions in recent years. Recently, deep learning-based algorithms have been successfully applied in the field of biomedical image segmentation.
Methods: In this study, an approach is proposed to automatically segment choroid layers associated with diabetic retinopathy and pachychoroid in optovue optical coherence tomography. For this purpose, U-Net architecture was used to segment BM and sclera boundaries. U-Net is the Fully Convolutional Network and modified in a way that it yields better segmentation in medical imaging and it contains three parts, namely down sampling path, Bottleneck and up sampling path. A new loss function with combination of four losses such as dice loss coefficient (DL), weighted categorical cross entropy (WCCE), Total Variation (TV) and Tversky is proposed. Subsequently, two class and three class segmentation is carried out.
Results: Three-class segmentation showed better performance compared to the two-class segmentation. Experimental results show a dice coefficient of 0.983 and 0.984 using three-class segmentation with combination of dice loss, WCCE and Tversky in diabetic retinopathy and pachychoroid patients respectively. Furthermore, unsigned error for diabetic retinopathy images is reported 0.025 and 0.29 mm for RPE and choroid layer respectively. For pachychoroid images, unsigned error is achieved 0.173 and 0.339 for RPE and choroid layer respectively.
Conclusion: By using deep learning algorithms such as U-Net, an improvement is achieved in choroid layer segmentation.
