Automatic Localization and Identification of Thoracic Diseases from Chest X-rays with Deep Learning

There are numerous difficulties in using deep learning to automatically locate and identify diseases in chest X-rays (CXR). The most prevailing two are the lack of labeled data of disease locations and poor model transferability between different datasets. This study aims to tackle these problems.

We built a new form of bounding box dataset and developed a two-stage model for disease localization and identification of CXRs based on deep learning. The dataset marks anomalous regions in CXRs but not the corresponding diseases, different from all previous datasets. The advantages of this design are reduced labor of annotation and fewer possible errors associated with image labeling. The two-stage model combines the robustness of the region proposal network, feature pyramid network, and multi-instance learning techniques. We trained and validated our model with the new bounding box dataset and the CheXpert dataset. Then, we tested its classification and localization performance on an external dataset, which is the official split test set of ChestX-ray14.

For classification result, the mean area under the receiver operating characteristic curve (AUC) metrics of our model on the CheXpert validation dataset was 0.912, which was 0.021, superior to the baseline model. The mean AUC of our model on an external testing set was 0.784, whereas the state-of-the-art model got 0.773. The localization results showed comparable performance to the stateof- the-art models.

Our model exhibits a good transferability between datasets. The new bounding box dataset is proven to be useful and shows an alternative technique for compiling disease localization datasets.