Visual Interpretability of Deep Learning Models in Glaucoma Detection Using Color Fundus Images
Journal
Proceedings - 2022 IET International Conference on Engineering Technologies and Applications, IET-ICETA 2022
ISBN
9781665491389
Date Issued
2022-01-01
Author(s)
Lu, Da Wen
Hsu, Wei Wen
Huang, Yu Chuan
Lee, Lung Chi
Guo, Jing Ming
Hsiao, Yu Ting
Lin, I. Hung
Chang, Yao Chung
Abstract
With the advancement of deep learning in many computer vision applications, the computer-aided diagnosis (CAD) systems were developed using the deep learning approaches to assist physicians in clinical diagnosis as a second opinion. Many studies have reported plausible performance of the deep learning models on glaucoma detection simply based on fundus images. However, the mechanism of the deep learning models in detecting glaucoma using color fundus is still not clear. Therefore, the deep features extracted by the deep learning models were discussed in this study to provide the visual interpretability for the frameworks of deep convolutional neural network (DCNN) in glaucoma assessment. In our experiments, 986 fundus photographs were collected from National Taiwan University Hospital Hsin-Chu Branch, categorized into two groups: the experimental group of 512 glaucomatous cases with ganglion cell complex (GCC) impairment and the control group of 474 non-glaucomatous cases with normal GCC thickness. The experimental results show the deep learning models mainly focus on the areas of optic nerve head (ONH) for the diagnosis of glaucoma, which is accordant to the clinical rules in glaucoma assessment. Surprisingly, the DCNN models can still achieved high prediction accuracy in detecting the glaucomatous cases with the cropped images of macular areas only. Several cases show the model's focus areas correspond to the region with GCC impairment. The results imply the deep learning models can detect the morphologically detailed alterations in fundus photographs, which may be beyond experts' visualization.
Subjects
deep convolutional neural network | fundus images | glaucoma detection | visual interpretability
SDGs
Type
conference paper