Exploring Meibomian Gland Dysfunction Grading: A Comparison of Machine Learning and XAI Approaches

This research investigates the effectiveness of employing a machine learning methodology and an Explainable Artificial Intelligence (XAI) tool for the assessment of Meibomian Gland Dysfunction (MGD). Utilizing ResNet34, a convolutional neural network (CNN), trained on a diverse dataset comprising images depicting varying degrees of MGD severity, the machine learning approach endeavors to classify MGD severity into four distinct grades (0, 1, 2, 3) by leveraging extracted characteristic features. Findings reveal comparable accuracy between ResNet34 and the XAI tool, with ResNet34 achieving a slightly lower accuracy of 99.1% in contrast to the XAI tool's 99.4%. However, ResNet34 demonstrates marginally higher precision and F1-score at 98.2% and 98.45%, respectively, suggesting a nuanced advantage in precision-recall equilibrium. Conversely, the XAI tool maintains a commendable precision of 97.2% and an Fl-score of 97.94%, alongside a recall of 98.7%, underscoring its effectiveness in discerning positive instances. Noteworthy is that ResNet34 operates as a black-box model, providing limited interpretability, while the XAI tool emphasizes transparency by furnishing clinicians with understandable MGD grading elucidations derived from meibomian gland segmentation and atrophy analysis. This transparency enhances trust and comprehension, which are integral for clinical acceptance and decision-making. Furthermore, the utilization of XAI heatmap generation facilitates data quality enhancement, as evidenced by instances of erroneous artifact identification, emphasizing the imperative nature of artifact elimination for the refinement of MGD grading procedures.