Abstract
Purpose :
The overall aim of the study is to explore the capability of detection using algorithms trained in Automated Deep Learning (AutoML) of graft failure from a proprietary dataset of post-keratoplasty patients from a case series published in the literature.
Methods :
Observational cross-sectional study, for which Automated Deep Learning 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, 10% validation).
Results :
The metrics obtained in the cloud-based and edge 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 average accuracy of 0.995, with a Specificity of 95.8% and a Sensitivity of 95.8%. For the label "Graft Failure" the algorithm presented a Specificity of 92.3% and a Sensitivity of 100%. Other metrics such as F1-score, AUC, confusion matrix and activation map development were contemplated.
Conclusions :
The initial results of our study indicate that it is possible to train algorithms in an automated fashion for the detection of Graft Failure in patients undergoing penetrating keratoplasty surgery. These algorithms, especially the edge type, are very lightweight tools and 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 this condition.
The metrics obtained in our training process are very good, but since it is a relatively small dataset, it is possible that we have some tendency to overfitting.
The use of Automated Machine Learning opens the possibility of working in the field of artificial intelligence by computer vision to professionals with little experience of programming and is a growing field.
In the next stages of our research line we will seek to expand the working dataset, train algorithms in a traditional way using the Tensorflow library and open ourselves to collaborations to increase the impact of our work.
This abstract was presented at the 2022 ARVO Annual Meeting, held in Denver, CO, May 1-4, 2022, and virtually.