Recognizing cardiac magnetic resonance acquisition acquisition planes usingfinetuned convolutional neural networks

In this paper we propose a convolutional neural network-based method to automatically wrangle missing or noisy cardiac acquisition plane information from magnetic resonance (MR) images. This is an important building block to organise and filter large collections of cardiac data prior to analysis. In addition it allows us to merge studies from multiple centers, to perform smarter image filtering, to select the most appropriate image processing algorithm, and to enhance visualisation of cardiac datasets in content based image retrieval. We propose to use a finetuned convolutional neural network initially trained on a large natural image recognition dataset (Imagenet ILSVRC2012) to learn feature representations for better recognize cardiac views prediction and contrast this to a previously introduced method using classification forests and features learned from an augmented set of image miniatures. We validated this algorithm on two different cardiac studies with 200 patients and 15 healthy volunteers respectively. Our new approach significantly improves the state of the art of image-based cardiac view recognition (97.66% F1 score). Despite the large number of the network’s parameters, the algorithm does not overfit and performs quite well on another independent cardiac study.