Abstract
Purpose :
Predicting progressive diseases in medical practice provides useful information to both doctors and patients. Recently, the sharing of information between doctors and patients is emphasized in order to fully understand the disease state and treatment. Thus, presenting future disease states to patients is effective to understand the treatment effect in follow-up. Particularly in ophthalmology, presenting progressive images is one of the most effective ways.
This study proposes a way to generate images with an arbitrary disease added to the fundus image of a patient, and further examines the technology to predict progressive diseases.
Methods :
The first approach of this method is to obtain a vascular map, a unique feature of a patient, from the fundus image and generate a disease image based on the map (Fig.1). This enables generation of an image with an arbitrary disease added to the fundus. Convolutional Neural Network (CNN) is used to generate the vascular map and the fundus image. Each CNN is trained based on Generative Advisory Network (GAN) used to generate images.
The second approach is to generate an image that predicts a progressive disease with a future disease map based on the diseased area detected from the fundus (Fig. 2).
The future disease map is obtained by extracting a target disease from the fundus image and generating the extracted disease area into a statistically predicted form. Also, an image predicting the form of the disease area after treatment can be generated.
Results :
For evaluation of generating a disease image, how accurately the generated image matched the real disease image was investigated, and subjective evaluation was conducted. In the generated disease image, the vascular part was 83.7% and the disease area was 95.7%. It was also confirmed that the generated disease image conformed at a high level to the real disease image by subjective evaluation.
Conclusions :
We generated disease images identifying a disease from the fundus image and the predicted images of how the disease will progress, and examined possibilities for application in predicting progressive diseases.
As a result, a disease image conforming to a real disease is generated. This suggests new possibilities for applications in predicting progressive diseases.
This abstract was presented at the 2019 ARVO Annual Meeting, held in Vancouver, Canada, April 28 - May 2, 2019.