Rethinking Retinal Image Quality: Treating Quality Threshold as a Tunable Hyperparameter

Abstract

Assuming the robustness of a deep learning model to suboptimal images is a key consideration, we asked if there was any value in including training images of poor quality. In particular, should we treat the (quality) threshold at which a training image is either included or excluded as a tunable hyperparameter? To that end, we systematically examined the effect of including training images of varying quality on the test performance of a DL model in classifying the severity of diabetic retinopathy. We found that there was a unique combination of (categorical) quality labels or a Goldilocks (continuous) quality score that gave rise to optimal test performance on either high-quality or suboptimal images. The model trained exclusively on high-quality images yielded worse performance in all test scenarios than that trained on the optimally tuned training set which included images with some level of degradation.

Retinal Changes From Hyperopia to Myopia: Not All Diopters Are Created Equal

“…In addition to this explanation, we cannot rule out the possibility that some of this association was due to selection bias inadvertently introduced by the exclusion of poor-quality images (e.g., perhaps centrally underexposed images with more negative SER were more likely to be rejected than similarly underexposed images with less negative SER, considering that the DL model that we used for image quality assessment might be more likely to reject “pathologic”-looking fundus photographs). 68 …”