Impact of gold-standard label errors on evaluating performance of deep learning models in diabetic retinopathy screening : nationwide real-world validation study

Abstract

Background: For medical artificial intelligence (AI) training and validation, human expert labels are considered the gold standard that represents the correct answers or desired outputs for a given data set. These labels serve as a reference or benchmark against which the model’s predictions are compared. Objective: This study aimed to assess the accuracy of a custom deep learning (DL) algorithm on classifying diabetic retinopathy (DR) and further demonstrate how label errors may contribute to this assessment in a nationwide DR-screening program. Methods: Fundus photographs from the Lifeline Express, a nationwide DR-screening program, were analyzed to identify the presence of referable DR using both (1) manual grading by National Health Service England–certificated graders and (2) a DL-based DR-screening algorithm with validated good lab performance. To assess the accuracy of labels, a random sample of images with disagreement between the DL algorithm and the labels was adjudicated by ophthalmologists who were masked to the previous grading results. The error rates of labels in this sample were then used to correct the number of negative and positive cases in the entire data set, serving as postcorrection labels. The DL algorithm’s performance was evaluated against both pre- and postcorrection labels. Results: The analysis included 736,083 images from 237,824 participants. The DL algorithm exhibited a gap between the real-world performance and the lab-reported performance in this nationwide data set, with a sensitivity increase of 12.5% (from 79.6% to 92.5%, P<.001) and a specificity increase of 6.9% (from 91.6% to 98.5%, P<.001). In the random sample, 63.6% (560/880) of negative images and 5.2% (140/2710) of positive images were misclassified in the precorrection human labels. High myopia was the primary reason for misclassifying non-DR images as referable DR images, while laser spots were predominantly responsible for misclassified referable cases. The estimated label error rate for the entire data set was 1.2%. The label correction was estimated to bring about a 12.5% enhancement in the estimated sensitivity of the DL algorithm (P<.001). Conclusions: Label errors based on human image grading, although in a small percentage, can significantly affect the performance evaluation of DL algorithms in real-world DR screening.