Abstract
Purpose :
As trachoma is eliminated worldwide, the number of adequately trained field graders for population-based prevalence surveys becomes more difficult to maintain. Ophthalmic photography and artificial intelligence (AI) have been cited as potential solutions for follicular trachoma (TF) surveillance. This applied research analyzed images collected during an Image Capture and Processing System (ICAPS) and AI validation study to systematically prioritize areas of improvement.
Methods :
Two independent graders scored a subset of 317 images from the ICAPS and AI validation study for a number of predefined image quality faults to explore false positivity, a particular challenge for the pilot AI. The subset included all 67 false positive images, 100 true negative images, 100 images deemed ungradable by an expert, and 50 images where the expert photograder and skilled field grader disagreed on the TF status. Univariable and multivariable logistic regression models were created to explore the association between gradeability, AI misclassification, and image quality faults. A prototype app was then built for the Android operating system and field tested in Tanzania.
Results :
In the 317 images, the quality faults of “blur” and “under-zoom” were more likely to occur together when controlling for “glare/lighting issues,” “uncentered” and “poor eversion/blanching” (OR = 4.094, 95% CI [2.372, 7.068]). Similarly, “under-zoom” and “uncentered” were more likely to occur together when controlling for the same variables (OR = 5.396, 95% CI [3.1, 9.395]). “Under-zoom” increased the odds of AI misclassification (OR = 2.7, 95% CI [1.352, 5.392]), while “poor eversion/blanching” decreased the odds of AI misclassification (OR = 0.334, 95% CI [0.174, 0.644]). Qualitative assessment of field images using the prototype Android app revealed better image quality.
Conclusions :
Understanding that “blur,” “under-zoom,” and “uncentered” were more likely to occur together, and that “under-zoom” had the greatest negative effect on AI classification, is helpful for targeting efficient improvements in future ICAPS. Standardization of the tarsal plate in the images may significantly improve image quality. Further refinements to the next generation camera system are planned.
This abstract was presented at the 2023 ARVO Annual Meeting, held in New Orleans, LA, April 23-27, 2023.