Hi! We used Vertex AI AutoML for video classification to classify echocardiograms of patients with reduced, mid-range and preserved left ventricular ejection fraction (a measure of the heart's function). Great success at first attempts with ROC AUC score around 95% and accuracy around 85% . However, at submission for publication, the reviewers complain about the opacity of AutoML and demand details about the classification models. We couldn’t find these details in the Vertex AI documentation. We would like to know which pre-trained CNN(s) are used (a 3D-Resnet? Others?), on which video dataset(s) they are pre-trained (Kinetic 400? Others?), and which hyperparameters are fine-tuned. We would be extremely grateful if some member of the Google AI team could provide this information. Details about our research with Vertex AI AutoML can be found here in Github.
LinkedIn
Twitter
