Use of a deep learning image quality model to evaluate impact of clinically relevant forms of degradation on the glaucoma gradability of fundus images

Abstract

Purpose : Poor image quality can adversely affect the ability of DL algorithms to detect glaucoma. This study evaluates the effect of several common types of image degradation on a DL model that automates quality assessment of fundus photographs.

Methods : 2,815 fundus images from the Diagnostic Innovations in Glaucoma Study and African Descent and Glaucoma Evaluation Study were used to develop a DL model to automate quality assessment. This model was tested on 11,350 photographs from the Ocular Hypertension Treatment Study. Several ablative operations were separately implemented to degrade high-quality images (QS=0) into low quality ones (QS≥0.1). Gaussian blur, brightness (bias) increases and decreases, contrast enhancement factor (gain) decreases, and Gaussian-distributed additive noise were applied. In order to determine the range of degradation values, images were degraded until quality scores peaked at a maximum quality value. For each type of ablation, the maximum degree of degradation was set as the value for which the resulting quality score first surpassed 95% of this peak poor quality value. Within this range, each image was incrementally degraded to create 100 progressively ablated photographs for that ablation form.

Results : While all degradations consistently yielded poor quality scores (Figure 1), each type of ablation yielded different degradation-QS profiles. Noise was the strongest determinant of quality, followed by contrast and focus. The model was more limited in recognizing the extent of degradation for both increased and decreased brightness (max QS = 0.45 and 0.76, respectively). The model appeared especially sensitive to the Gaussian noise ablation type; random distribution variance of 0.0075 and 0.0125 proved sufficient to degrade an image to a poor quality score.

Conclusions : A DL model to assess fundus photography quality can offer insights into factors influencing DL models for detection of eye disease. Evaluation of the performance of a glaucomatous optic neuropathy model on variably degraded datasets in conjunction with grading by human experts may validate or refute these insights.

This abstract was presented at the 2022 ARVO Annual Meeting, held in Denver, CO, May 1-4, 2022, and virtually.

 

Figure 1. Degradations at key QS thresholds. 1) No degradation 2) First degradation exceeding low quality threshold 3) Peak degradation. A) Gaussian blur B) Increased brightness C) Decreased brightness D) Decreased contrast E) Increased noise.