Binary and Random Inputs to Rapidly Identify Overfitting of Deep Neural Networks Trained to Output Ultrasound Images

Abstract

We developed a novel method to detect overfitting of deep neural networks trained to create ultrasound images. This method only requires the network architecture and trained weights, and does not require loss function monitoring during an otherwise time-consuming training process. Specifically, two binary images and an image of Gaussian random noise were used as inputs to three neural networks submitted to the Challenge on Ultrasound Beamforming with Deep Learning (CUBDL). Comparing the network-created images to the ground truth immediately revealed an overfit to the data used to train one of the three networks, indicating the promise of our method to detect overfitting without requiring lengthy network retraining or the collection of additional test data. This approach holds promise for regulatory oversight of DNNs intended to be deployed on patient data.