advertisement
Editors Selection IGR 16-1
Artificial Intelligence: AI in Glaucoma Screening
Comment by Minguang He on:
109556 A generalizable deep learning regression model for automated glaucoma screening from fundus images, Hemelings R; Elen B; Schuster AK et al., NPJ digital medicine, 2023; 6: 112
This study marks an important advancement in the field of glaucoma detection using Convolutional Neural Networks (CNNs). The robustness of these networks was tested across thirteen external datasets, including two large population cohorts and eleven publicly accessible datasets. This approach represents a significant step in achieving generalizability in glaucoma detection. The research confirms the effectiveness of CNNs in detecting glaucoma, making it a comprehensive analysis of generalizability to date.
A key aspect of this study is the use of a regression-based model in CNNs, which differs from traditional binary classification methods. By assessing a range of disease severities, such as the vertical cup disc ratio, rather than simply classifying conditions as 'with or without glaucoma', the model is able to capture more detailed variations in the disease. This, coupled with soft labeling techniques, enhances the model's ability to generalize and speeds up its convergence.
The continuous regression scoring approach risks unbalanced learning or overfitting, particularly if there is an insufficient variety of training samplesDespite challenges, such as the overprediction of risk scores in CNNs using sigmoid activation, the study presents a consistent saliency pattern, especially in the infero- and supero-temporal regions of the eye. This pattern was observed across a test set of over 4,000 fundus images. However, the study may be limited by its focus on glaucoma-specific cases and does not account for a wider range of comorbidities, for example the use of population-based data for training resulted in an underrepresentation of severe glaucoma cases, and it did not include other co-existing eye diseases. From a technical perspective, the continuous regression scoring approach risks unbalanced learning or overfitting, particularly if there is an insufficient variety of training samples. Clinically, implementing any glaucoma detection algorithm in practice would necessitate a prospective trial to determine its effectiveness, accuracy, and practicality, as well as the need for regulatory approval.
Comments
The comment section on the IGR website is restricted to WGA#One members only. Please log-in through your WGA#One account to continue.Log-in through WGA#One
