New Ordinal Regression and Endurance Test for Scheie Classification in Fundus Images

Abstract

Purpose : The Scheie Classification is commonly employed in categorizing hypertensive retinopathy, yet the prevalence of high-grade cases is low, resulting in a significant skew that impedes machine learning. In this study, we comprehensively tested five types of ordinal regression methodologies along with our novel approach against the Scheie Classification's arteriolar sclerosis. Additionally, we evaluated the resilience of inferential reasoning using the Citadel Lens, an AI-based automatic endurance test.

Methods : Utilizing a dataset of 320,000 fundus photographs collected over the past decade by the Health Care Center at Jichi Medical University, five methodologies, namely, Beckham, Pan, Niu, Diaz, and Liu were trained over 150 epochs using the Adam optimizer. These methods were evaluated using the Mean Absolute Error (MAE) metric. Our proposed technique enhances the loss function by integrating multiple cross-entropy losses, incorporating the mean and variance, employing Kullback-Leibler divergence, and utilizing optimal transport. Additionally, we refined the network output by constraining it to a Poisson distribution, aiming to improve prediction accuracy and reliability.

Results : In an assessment of retinal image grading distributions, the dataset comprised 277,800 images at grade 0, 50,005 at grade 1, 13,442 at grade 2, 312 at grade 3, and 10 at grade 4. Assuming an equal interval scale, the calculated Gini coefficient was 0.91, indicating a high concentration of lower grades. MAE for the traditional ordinally regression methodologies yielded results as follows: 0.22, 0.15, 0.10, 0.14, 0.17, 0.11, and 0.09 for Beckham, Pan, Niu, Diaz, Liu, the baseline method not considering ordinal scales, and our method, respectively. The Citadel Lens outcomes for our model demonstrated an Explainability of 84%, Model Bias at 51%, Error Analysis at 59%, Camera Robustness at 76%, Environmental Robustness at 70%, and Position Robustness at 82%.

Conclusions : In the classification of arteriosclerotic observations into five categories, which displayed a highly skewed distribution with a Gini coefficient of 0.91, our novel method achieved the lowest MAE of 0.09. Furthermore, our model demonstrated a relatively low model bias and a comparatively high level of explainability.

This abstract was presented at the 2024 ARVO Annual Meeting, held in Seattle, WA, May 5-9, 2024.

 

Neural Network Architecture