Abstract
Purpose :
The irreversible blinding of glaucoma makes it a critical need for the patient for an early and accurate diagnosis of glaucoma. While the state-of-art deep neural network is developed rapidly, we propose a new network model and make it applicable to the computer-aided diagnosis based on the data of Heidelberg Retina Tomograph (HRT) of glaucoma with weak labels, which includes HRT images of two modalities: the reflection intensity of light and the topographic value.
Methods :
We propose a novel deep network model to make full use of the HRT data with weak labels, which represent the characteristics of two types of images: 1. the reflection intensity of light, 2. the topographic value. The design of our network model can solve the ambiguity of the two types of images when they are input together and maintain the subtle differences when the two types of images are input separately. In addition, the visualization technique is used for the proposed network of normal subjects versus glaucoma ones, which could show the supported evidence for the decision of the network.
Results :
After training the proposed deep neural network model for the diagnosis of glaucoma using the HRT images with weak labels, the independent test experiment made 94.0% of the area under the ROC curve (AUC), which means at least a 5% increase compared with the previous results. More importantly, the sensitivity is 91.2% when the specificity is 0.85 and is 78.3% when the specificity is 0.95, which implies that our method may potentially be used for the clinical screening as an indicator for the early diagnosis of glaucoma.
Conclusions :
Our work can change the diagnositic mode of glaucoma using HRT data. Independent test results show that the proposed network model can contribute to an accurate diagnosis of glaucoma (especially in the early stage) with HRT. An early and accurate diagnosis of glaucoma may lead to early and precise treatment for the patients with glaucoma, and may result in an increase in their quality of life.
This is an abstract that was submitted for the 2018 ARVO Annual Meeting, held in Honolulu, Hawaii, April 29 - May 3, 2018.