Jinho Lee1,2, Youngwoo Kim3,4, Jong Hyo Kim3,4,5, Ki Ho Park1,2. 1. Department of Ophthalmology. 2. Department of Ophthalmology, Seoul National University Hospital, Seoul. 3. Interdisciplinary Program of Radiation Applied Life Science, Seoul National University College of Medicine. 4. Center for Medical-IT Convergence Technology Research, Advanced Institutes of Convergence Technology. 5. Graduate School of Convergence Science and Technology, Seoul National University, Gwanggyo, Korea.
Abstract
PRéCIS:: The novel proposed algorithm using deep learning classifier and polar transformation technique can be an economical as well as an effective tool for early detection of glaucomatous RNFL defect. PURPOSE: The main purpose of this study was to develop novel software to determine whether there is a retinal nerve fiber layer (RNFL) defect in a given fundus image using deep learning classifier and, if there is, where it presents. MATERIALS AND METHODS: In the deep learning classifier, the bottleneck features were extracted, followed by application of the softmax classifier, which outputted the glaucoma probability. For localization of RNFL defect, an image processing algorithm was implemented as follows: (1) the given image was normalized to enhance the contrast; (2) the region of interest (ROI) was set as the circumferential area surrounding the optic disc (internal diameter: 2 disc diameters, external diameter: 3 disc diameter), and converted to a polar image; (3) blood vessels were removed and the average curvatures were calculated. If the local maximum curvature was greater than the cut-off value, the sector was considered to be an RNFL defect.The images of 100 normal healthy controls and 100 open-angle glaucoma patients were enrolled. Maximum curvatures and area under receiver operating characteristic curve were compared to determine the diagnostic validity. RESULTS: There were no significant differences in age or sex (P=0.275, P=0.479, respectively) between the 2 groups. In the glaucoma group, the mean deviation was -4.9±5.4 dB. There was a significant difference of maximum curvature (14.37±5.13 in control group, 20.67±10.56 in glaucoma group, P<0.001). Area under receiver operating characteristic curve was 0.939 in deep learning classifier and 0.711 in maximum curvature. CONCLUSIONS: The proposed software can be an effective tool for automated detection of RNFL defect.
PRéCIS:: The novel proposed algorithm using deep learning classifier and polar transformation technique can be an economical as well as an effective tool for early detection of glaucomatous RNFL defect. PURPOSE: The main purpose of this study was to develop novel software to determine whether there is a retinal nerve fiber layer (RNFL) defect in a given fundus image using deep learning classifier and, if there is, where it presents. MATERIALS AND METHODS: In the deep learning classifier, the bottleneck features were extracted, followed by application of the softmax classifier, which outputted the glaucoma probability. For localization of RNFL defect, an image processing algorithm was implemented as follows: (1) the given image was normalized to enhance the contrast; (2) the region of interest (ROI) was set as the circumferential area surrounding the optic disc (internal diameter: 2 disc diameters, external diameter: 3 disc diameter), and converted to a polar image; (3) blood vessels were removed and the average curvatures were calculated. If the local maximum curvature was greater than the cut-off value, the sector was considered to be an RNFL defect.The images of 100 normal healthy controls and 100 open-angle glaucomapatients were enrolled. Maximum curvatures and area under receiver operating characteristic curve were compared to determine the diagnostic validity. RESULTS: There were no significant differences in age or sex (P=0.275, P=0.479, respectively) between the 2 groups. In the glaucoma group, the mean deviation was -4.9±5.4 dB. There was a significant difference of maximum curvature (14.37±5.13 in control group, 20.67±10.56 in glaucoma group, P<0.001). Area under receiver operating characteristic curve was 0.939 in deep learning classifier and 0.711 in maximum curvature. CONCLUSIONS: The proposed software can be an effective tool for automated detection of RNFL defect.