Construction of a Computational Framework to Automatically Interpret Chest X-rays and Diagnose Pneumonia

This study aimed to systematically diagnose pneumonia directly from paediatric chest X-ray images using a computational framework. The project research goals were 1) to establish a high-quality dataset of pneumonia labelled X-ray images, 2) to extend existing deep learning architectures for pneumonia diagnoses, and 3) to construct a computational pipeline, enabling members of the broader community to interface with the computational models to diagnose pneumonia from X-ray images. The paediatric chest x-ray images were derived from 1) a WHO-supported surveillance study at Dhaka Shishu Hospital (DSH), from 2) a community site at Kumudini Women's Medical College (KWMCH), and from 3) the WHO Chest Radiography in Epidemiological Studies (WHO CRES) working group. The images were interpreted by at least two trained clinicians / radiologists for the presence of 1) primary end-point pneumonia (PEP), 2) other lung infiltrates, and/or 3) pleural fluid in either the left or right lung, for a total of six possible binary outcomes, which will henceforth be called 'labels'. A third reader resolved discordant PEP labels found between the first and second readers. An X-ray image were included in this study if 1) the age of the child for whom the X-ray was performed was <=59 months, 2) the X-ray was performed in one of the two study hospitals or from the WHO CRES reference image set, and 3) the image captured the lung area. Any Image that was marked 'uninterpretable' (features of the images were not interpretable with respect to presence or absence of PEP) by two readers were excluded. The deposited datasets contain the resulting labels from the multiple readers for each dataset described above.