Cardiomegaly, radiographically measured by a cardiothoracic ratio of greater than 50%, is a common symptom associated with congestive heart failure [1]. Given the large size of the heart relative to the view covered by the thoracic X-ray, its detection should be a relatively easy task for both humans and neural networks. Cardiomegaly detection thus serves as a useful baseline for the comparison neural network architectures.

In 2017, Wang et al., in collaboration with the NIH Clinical Center, released a 110,000+ image dataset, Chest-Xray14, with NLP-generated annotations for 14 conditions found on chest radiographs, including cardiomegaly [2]. The AUC of the model Wang et al. utilized for cardiomegaly detection was of 0.807. Improving on that result, the DenseNet-based CheXNet model by Rajpurkkar et al. obtained an AUC of 0.92 on cardiomegaly [3].

We want to evaluate the performance and efficacy of state-of-the-art deep convolutional neural networks (DCNNs) for automatic detection of cardiomegaly on chest radiographs, and compare their performance against older convolutional architectures.