Number of DiDA response areas to fundus images and confidence in estimated age

Abstract

Machine learning has enabled accurate estimation of true age from fundus images. However, it is unclear which image features are important, and which parts of the image are most clinically relevant, for age estimation. Although methods such as Grad-CAM and DiDA can be used to interpret where the machine learning model looks for inferences, most studies have focused on object detection and classification, with few investigating regression problems. In this paper, we apply DiDA to two different models for age estimation from fundus images and examines the relationship between the sizes of the responding regions and the errors in the estimated age. In addition to counting the number of reacted pixels, we divided the fundus images into three regions, and the counted reacted pixels in each region to decide whether the region was reacted or not. The results show that a smaller error between the actual and estimated age, leads to a greater number of reacted areas and the more accurate the estimated age. These results suggest that the DiDA algorithm can be used to extract the confidence level of the age estimation model.