Detection of graft failure in post-keratoplasty patients by automated deep learning

Detection of graft failure of post-penetrating keratoplasty (PKP) patients from the proprietary dataset using algorithms trained in Automated Deep Learning (AutoML).

This was an observational cross-sectional study, for which AutoML algorithms were trained following the success/failure labeling strategy based on clinical notes, on a cohort corresponding to 220 images of post-keratoplasty anterior pole eyes. Once the image quality criteria were analyzed and the dataset was pseudo-anonymized, it was transferred to the Google Cloud Platform, where using the Vertex AI-AutoML API, cloud- and edge-based algorithms were trained, following expert recommendations on dataset splitting (80% training, 10% test, and 10% validation).

The metrics obtained in the cloud-based and edge-based models have been similar, but we chose to analyze the edge model as it is an exportable model, lighter and cheaper to train. The initial results of the model presented an accuracy of 95.83%, with a specificity of 91.67% and a sensitivity of 100%, obtaining an F1_SCORE of 95.996% and a precision of 92.30%. Other metrics, such as the area under the curve, confusion matrix, and activation map development, were contemplated.

Initial results indicate the possibility of training algorithms in an automated fashion for the detection of graft failure in patients who underwent PKP. These algorithms are very lightweight tools easily integrated into mobile or desktop applications, potentially allowing every corneal transplant patient to have access to the best knowledge to enable the correct and timely diagnosis and treatment of graft failure. Although the results were good, because of the relatively small dataset, it is possible the data have some tendency to overfitting. AutoML opens the possibility of working in the field of artificial intelligence by computer vision to professionals with little experience and knowledge of programming.