Abstract
Categorizing medical samples is a difficult and time-consuming task that directly impacts patient outcomes. Recent technological advancements may hold the key to improving medical professionals' diagnostic accuracy. One of these advancements is EchoNet-Dynamic, a convolutional neural network that segments echocardiograms-ultrasound videos of the heart-producing a red overlay onto the left ventricle, the area of the heart relevant to diagnosis. We investigated the potential for EchoNet-Dynamic's segmentation to aid naïve non-clinician humans and pigeons in their diagnosis of cardiac function. Humans were trained to categorize either segmented or non-segmented echocardiograms as depicting normal or abnormal heart function. Then, roughly half of the subjects in each group were tested with videos of the opposite type they were trained with. We found that more humans trained with segmented videos adequately learned the task than those trained with non-segmented videos; they also learned more quickly, exhibited higher accuracies at the end of training, and reliably generalized to non-segmented videos during testing. Despite these apparent benefits, there was no general improvement in the accuracy of humans trained with non-segmented videos when testing with segmented videos. Pigeons, trained with segmented videos, successfully learned the task. However, unlike humans, they failed to generalize their learning to non-segmented videos, even after a fading procedure was employed. We conclude that EchoNet-Dynamic's segmentation is an effective visual aid that enhances learning and enables reliable transfer to non-segmented videos for humans, and provides a means of learning what otherwise might have been an incredibly difficult task for pigeons.