TY - GEN
T1 - 3D-CNN for Glaucoma Detection Using Optical Coherence Tomography
AU - George, Yasmeen
AU - Antony, Bhavna
AU - Ishikawa, Hiroshi
AU - Wollstein, Gadi
AU - Schuman, Joel
AU - Garnavi, Rahil
N1 - Publisher Copyright:
© Springer Nature Switzerland AG 2019.
PY - 2019
Y1 - 2019
N2 - The large size of raw 3D optical coherence tomography (OCT) volumes poses challenges for deep learning methods as it cannot be accommodated on a single GPU in its original resolution. The direct analysis of these volumes however, provides advantages such as circumventing the need for the segmentation of retinal structures. Previously, a deep learning (DL) approach was proposed for the detection of glaucoma directly from 3D OCT volumes, where the volumes were significantly downsampled first. In this paper, we propose an end-to-end DL model for the detection of glaucoma that doubles the number of input voxels of the previously proposed method, and also boasts an improved AUC = 0.973 over the results obtained using the previously proposed approach of AUC = 0.946. Furthermore, this paper also includes a quantitative analysis of the regions of the volume highlighted by grad-CAM visualization. Occlusion of these highlighted regions resulted in a drop in performance by 40%, indicating that the regions highlighted by gradient-weighted class activation maps (grad-CAM) are indeed crucial to the performance of the model.
AB - The large size of raw 3D optical coherence tomography (OCT) volumes poses challenges for deep learning methods as it cannot be accommodated on a single GPU in its original resolution. The direct analysis of these volumes however, provides advantages such as circumventing the need for the segmentation of retinal structures. Previously, a deep learning (DL) approach was proposed for the detection of glaucoma directly from 3D OCT volumes, where the volumes were significantly downsampled first. In this paper, we propose an end-to-end DL model for the detection of glaucoma that doubles the number of input voxels of the previously proposed method, and also boasts an improved AUC = 0.973 over the results obtained using the previously proposed approach of AUC = 0.946. Furthermore, this paper also includes a quantitative analysis of the regions of the volume highlighted by grad-CAM visualization. Occlusion of these highlighted regions resulted in a drop in performance by 40%, indicating that the regions highlighted by gradient-weighted class activation maps (grad-CAM) are indeed crucial to the performance of the model.
KW - 3D-CNN
KW - Glaucoma detection
KW - Gradient-weighted class activation maps
KW - Optical coherence tomography
KW - Visual explanations
UR - http://www.scopus.com/inward/record.url?scp=85075660036&partnerID=8YFLogxK
UR - http://www.scopus.com/inward/citedby.url?scp=85075660036&partnerID=8YFLogxK
U2 - 10.1007/978-3-030-32956-3_7
DO - 10.1007/978-3-030-32956-3_7
M3 - Conference contribution
AN - SCOPUS:85075660036
SN - 9783030329556
T3 - Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)
SP - 52
EP - 59
BT - Ophthalmic Medical Image Analysis - 6th International Workshop, OMIA 2019, Held in Conjunction with MICCAI 2019, Proceedings
A2 - Fu, Huazhu
A2 - Garvin, Mona K.
A2 - MacGillivray, Tom
A2 - Xu, Yanwu
A2 - Zheng, Yalin
PB - Springer
T2 - 6th International Workshop on Ophthalmic Medical Image Analysis, OMIA 2019, held in conjunction with the 22nd International Conference on Medical Imaging Computing and Computer-Assisted Intervention, MICCAI 2019
Y2 - 17 October 2019 through 17 October 2019
ER -