|
목적 : To compare the performance of different deep learning architectures to detect optic nerve head (ONH), macula, and cup-to-disc ratio (CDR) in fundus image.
방법 : 1726 fundus images from 868 normal and glaucoma patients were collected and of these images, 1380 (80%) / 346 (20%) images were used as neural network training / validation. Vertical CDR was determined by ONH parameters measured with Zeiss spectral domain Cirrus optical coherence tomography (OCT). Four different state-of-art deep learning architectures were trained: YOLOV3, YOLOV3-Tiny, Resnet50, Densenet201. To test performance, 3 parameters were calculated: detection time, the number of CDR matching, intersection of union (IoU) of the ONH and macula bounding box.
결과 : Mean detection time was 4.02 / 12.43 / 19.33 / 28.01 ms with GPU and was 266 / 1237 / 1725 / 2701 ms without GPU (YOLOV3-Tiny, Resnet50, Densenet201, YOLOV3, respectively). Mean IoU of the ONH was 82.8% / 81.5% / 75.3% / 0.6% and that of macula was 79.1% / 77.3% / 75.2% / 59.2% (YOLOV3, YOLOV3-Tiny, Resnet50, Densenet201, respectively). The number of predicted CDR exactly matching to ground truth was 67.9% / 67.5% / 66.4% / 11.4 % and and the number of nearly matched (CDR difference ±0.2) was 31.4% / 29.9% / 29.2% / 25.1% (YOLOV3, Resnet50, YOLOV3-Tiny, Densenet201, respectively.
결론 : YOLOV3-Tiny showed the fastest detection performance. With GPU assist, detection time was accelerated by about 100 times faster than CPU only. The best IoU of both ONH and macula was achieved by YOLOV3 but the difference between YOLOV3, YOLOV3-Tiny and Resnet50 was clinically negligible. The CDR prediction of YOLOV3 was best and its accuracy was 67.9% (exactly matched) and 31.4% (nearly matched).
* This research was supported by the Bio & Medical Technology Development Program of the National Research Foundation (NRF) funded by the Korean government (MSIT) (No. NRF-2018M3A9E8066253)
|