TY - GEN
T1 - Real-time retinal layer segmentation of adaptive optics optical coherence tomography angiography with deep learning
AU - Jian, Yifan
AU - Borkovkina, Svetlana
AU - Japongsori, Worawee
AU - Camino, Acner
AU - Sarunic, Marinko V.
N1 - Publisher Copyright:
© 2020 IEEE.
PY - 2020/9
Y1 - 2020/9
N2 - Real time rendering of en face optical coherence tomography (OCT) and OCT-angiography (OCTA) of arbitrary retinal layers in ophthalmic imaging sessions can be used to increase the yield rate of high-quality acquisitions, provide real-time feedback during image-guided surgeries and compensate aberrations in sensorless adaptive optics (AO) OCT and OCTA. However, real-time en face visualizations rely critically on the accurate segmentation of retinal layers in the three-dimensional OCT volumes. Here, we demonstrate a compact deep-learning architecture that segmented batches of OCT B-scans and produced the corresponding OCT and OCTA projections within only 41 ms. The short latency was possible due to a low complexity neural network structure, CNN compression using TensorRT, and the use of Tensor Cores on GPU hardware to accelerate the computation of convolutions. Inferencing of the original U-net was accelerated by 21 times without reducing the accuracy. To the best our knowledge, our work is the first demonstration of an ophthalmic imager with embedded artificial intelligence (AI) providing real-time feedback.
AB - Real time rendering of en face optical coherence tomography (OCT) and OCT-angiography (OCTA) of arbitrary retinal layers in ophthalmic imaging sessions can be used to increase the yield rate of high-quality acquisitions, provide real-time feedback during image-guided surgeries and compensate aberrations in sensorless adaptive optics (AO) OCT and OCTA. However, real-time en face visualizations rely critically on the accurate segmentation of retinal layers in the three-dimensional OCT volumes. Here, we demonstrate a compact deep-learning architecture that segmented batches of OCT B-scans and produced the corresponding OCT and OCTA projections within only 41 ms. The short latency was possible due to a low complexity neural network structure, CNN compression using TensorRT, and the use of Tensor Cores on GPU hardware to accelerate the computation of convolutions. Inferencing of the original U-net was accelerated by 21 times without reducing the accuracy. To the best our knowledge, our work is the first demonstration of an ophthalmic imager with embedded artificial intelligence (AI) providing real-time feedback.
KW - angiography
KW - deep learning
KW - ophthalmic imaging
KW - optical coherence tomography
KW - real time
KW - segmentation
UR - http://www.scopus.com/inward/record.url?scp=85097863521&partnerID=8YFLogxK
UR - http://www.scopus.com/inward/citedby.url?scp=85097863521&partnerID=8YFLogxK
U2 - 10.1109/IPC47351.2020.9252343
DO - 10.1109/IPC47351.2020.9252343
M3 - Conference contribution
AN - SCOPUS:85097863521
T3 - 2020 IEEE Photonics Conference, IPC 2020 - Proceedings
BT - 2020 IEEE Photonics Conference, IPC 2020 - Proceedings
PB - Institute of Electrical and Electronics Engineers Inc.
T2 - 2020 IEEE Photonics Conference, IPC 2020
Y2 - 28 September 2020 through 1 October 2020
ER -