Automated Detection of Pathological Myopia from Colour Fundus Photographs using Code-free Deep Learning
Abstract:
Introduction: The maculopathy observed in highly myopic eyes is intricate. Clinical diagnosis of this condition poses a significant workload and is subjective, particularly in Asian countries such as China. To streamline and expedite the classification of pathologic myopia (PM), a code-free deep learning (CFDL) model was created utilizing a novel platform, specifically the Huawei Cloud, which is accessible to regions with restricted availability of other commercial platforms. Subsequently, the performance of the CFDL model was assessed against a deep learning (DL) algorithm designed for screening PM lesions using color fundus photographs (CFPs).
Methods: This research involved analyzing 12,000 CFPs obtained from 1200 individuals in the PALM dataset. Out of these images, 10,800 were utilized for training purposes, while the remaining 1200 were allocated for validation. A CFDL algorithm was developed, validated, and tested on the Huawei Cloud platform to screen for PM in CFPs. Additionally, a custom DL model was trained using transfer learning and the EfficientNet-B8 architecture on the training dataset. The study compared the performance metrics, including the area under the receiver operating characteristic curve (AUROC), sensitivity, specificity, and accuracy, of the CFDL model with those of the bespoke DL model.
Results: In the validation dataset, the CFDL model exhibited AUROCs ranging from 0.95 to 0.98 for detecting PM and normal fundus, sensitivities between 92.90% and 93.10%, specificities from 93% to 97.60%, and an overall accuracy of 92.90%. Conversely, the custom traditional DL model achieved AUROCs of 0.94 to 0.96, sensitivities ranging from 91.2% to 92%, specificities between 92.5% and 94%, and an overall accuracy of 91.8%.
Conclusion: We created a CFDL model on the Huawei cloud platform to detect and screen for PM using CFPs. Our model demonstrated superior sensitivities, specificities, and consistent accuracies when compared to conventional DL methods.
