Abstract
Purpose :
Although numerous numbers of accurately annotated data as grand truth is necessary for deep learning, the preparation of grand truth can create a bottleneck. In this study, we prepared a small number of annotated corneal endothelial images in a early-stage Fuchs endothelial corneal dystrophy (FECD) model mouse, and tested the feasibility of semi-supervised learning to widen the indication of AI to late-stage FECD.
Methods :
Corneal endothelial images were obtained from FECD mouse-model eyes via contact specular microscopy. A trained model (AI 1) was created via the use of 28 manually annotated images (ground truth) of early-stage FECD mouse eyes. Supervised data 1 (n=250) of late-stage FECD was generated via AI 1, and AI 2 was generated via the use of supervised data 1. Then, supervised data 2 was generated by AI 2, and AI 3 was generated by using supervised data 2. Subsequently, those learning processes were repeated up to AI 12. Finally, AI was used to analyze the 25 test data images of late-stage FECD.
Results :
AI 1 was generated via the 28 annotated image data of early-stage FECD. The guttae area detected by AI 1 was strongly associated with ground truth in early-stage FECD (r=0.97, p=3.83×10-27), however, AI 1 underestimated the guttae area by a mean systematic error (i.e., between the guttae area detected by AI and ground truth) of -2.2 ± 1.3% (r=0.86, p=2.83×10-8) in late-stage FECD. After semi-supervised learning, systematic error tended to decrease throughout AI 2-9, though it increased slightly due to overestimation in AI 10-12. The mean systematic error of AI 9 was -0.1±0.9%, and the guttae area detected by AI 9 was strongly associated with the manually annotated test data of late-stage FECD (r=0.94, p=5.84×10-12).
Conclusions :
Semi-supervised learning by the use of limited numbers of annotated data of early-stage FECD enables the generation of a deep neural network for analyzing different disease phases of late-stage FECD. Our data suggests that semi-supervised learning can be applicable to multiple clinical settings; e.g., reduction of the time and cost of preparing annotated data and expanded indication (to very early or late stages) of a deep neural network.
This is a 2021 ARVO Annual Meeting abstract.