TY - JOUR
T1 - Joint Segmentation and Quantification of Chorioretinal Biomarkers in Optical Coherence Tomography Scans
T2 - A Deep Learning Approach
AU - Hassan, Bilal
AU - Qin, Shiyin
AU - Hassan, Taimur
AU - Ahmed, Ramsha
AU - Werghi, Naoufel
N1 - Funding Information:
Manuscript received February 16, 2021; revised April 10, 2021; accepted April 26, 2021. Date of publication May 6, 2021; date of current version May 19, 2021. This work was supported by a research fund from Khalifa University under Grant CIRA-2019-047. The Associate Editor coordinating the review process was Octavian Adrian Postolache. (Corresponding author: Taimur Hassan.) Bilal Hassan is with the School of Automation Science and Electrical Engineering, Beihang University (BUAA), Beijing 100191, China.
Publisher Copyright:
© 1963-2012 IEEE.
PY - 2021
Y1 - 2021
N2 - In ophthalmology, chorioretinal biomarkers (CRBMs) play a significant role in detecting, quantifying, and ameliorating the treatment of chronic eye conditions. Optical coherence tomography (OCT) imaging is primarily used for investigating various CRBMs and prompt intervention of retinal conditions. However, with extensive clinical applications and increasing prevalence of ocular diseases, the number of OCT scans obtained globally exceeds ophthalmologists' capacity to examine these in a meaningful manner. Instead, the emergence of deep learning provides a cost-effective and reliable alternative for automated analysis of scans, assisting ophthalmologists in clinical routines and research. This article presents a residual learning-based framework (RASP-Net) that integrates atrous spatial pyramid pooling, coherent preprocessing, and postprocessing mechanisms to achieve joint segmentation and quantification of 11 CRBMs. We used a total of 7000 annotated scans for training, validation, and testing purposes of RASP-Net. Moreover, a novel algorithm for 3-D macular profiles reconstruction is presented to give a more intuitive way for characterizing the CRBMs based on coarse contouring and quantification. The proposed framework is evaluated through several experiments using different performance metrics. The results presented in this study validate the optimal performance of RASP-Net in precise detection and segmentation of CRBMs, with mean balanced accuracy, intersection over union, and dice score values of 0.916, 0.634, and 0.776 respectively. The proposed RASP-Net model characterizes a wide range of CRBMs with fine-grained pixelwise segmentation, extraction, and quantification in the context of retinal pathologies. This proposed system can allow retina experts to monitor the improvement and deterioration of the underlying ocular conditions.
AB - In ophthalmology, chorioretinal biomarkers (CRBMs) play a significant role in detecting, quantifying, and ameliorating the treatment of chronic eye conditions. Optical coherence tomography (OCT) imaging is primarily used for investigating various CRBMs and prompt intervention of retinal conditions. However, with extensive clinical applications and increasing prevalence of ocular diseases, the number of OCT scans obtained globally exceeds ophthalmologists' capacity to examine these in a meaningful manner. Instead, the emergence of deep learning provides a cost-effective and reliable alternative for automated analysis of scans, assisting ophthalmologists in clinical routines and research. This article presents a residual learning-based framework (RASP-Net) that integrates atrous spatial pyramid pooling, coherent preprocessing, and postprocessing mechanisms to achieve joint segmentation and quantification of 11 CRBMs. We used a total of 7000 annotated scans for training, validation, and testing purposes of RASP-Net. Moreover, a novel algorithm for 3-D macular profiles reconstruction is presented to give a more intuitive way for characterizing the CRBMs based on coarse contouring and quantification. The proposed framework is evaluated through several experiments using different performance metrics. The results presented in this study validate the optimal performance of RASP-Net in precise detection and segmentation of CRBMs, with mean balanced accuracy, intersection over union, and dice score values of 0.916, 0.634, and 0.776 respectively. The proposed RASP-Net model characterizes a wide range of CRBMs with fine-grained pixelwise segmentation, extraction, and quantification in the context of retinal pathologies. This proposed system can allow retina experts to monitor the improvement and deterioration of the underlying ocular conditions.
KW - Chorioretinal biomarkers (CRBMs)
KW - deep learning
KW - lesions segmentation
KW - optical coherence tomography (OCT)
KW - quantification
KW - retinal image analysis
UR - http://www.scopus.com/inward/record.url?scp=85105850124&partnerID=8YFLogxK
U2 - 10.1109/TIM.2021.3077988
DO - 10.1109/TIM.2021.3077988
M3 - Article
AN - SCOPUS:85105850124
SN - 0018-9456
VL - 70
JO - IEEE Transactions on Instrumentation and Measurement
JF - IEEE Transactions on Instrumentation and Measurement
M1 - 9424575
ER -