Using Deep Learning for Pneumothorax Detection: Newly Published Research from UCSF Radiology

A collapse of the lung due to air in the chest, known as pneumothorax, can be a life- threatening emergency. Pneumothorax is often detected by chest X-ray, but delays in review of these images can lead to delay in diagnosis and treatment. The goal of research is to solve problems and improve clinical outcomes; faster detection and better care in this case. In radiology, machine learning models such as deep convolutional neural networks are being increasingly used for medical image analysis. An automated method of screening chest X-rays and prioritization of images that are suspected to show a pneumothorax for rapid review may result in earlier treatment of pneumothorax.

With this in mind, researchers at the UC San Francisco Department of Radiology and Biomedical Imaging, and the Center for Digital Health Innovation (CDHI) set out to develop computer algorithms that scan chest X-rays and flag images that are suspicious for containing a moderate or large pneumothorax. By training on a large set of both positive and negative chest X-rays, these algorithms “learned” to identify moderate and large-sized pneumothorax.

For this retrospective study, researchers created the training set of images by asking board-certified radiologists to label each image for the presence or absence of pneumothorax, as well as their estimate of pneumothorax size. In all, 13,292 frontal chest X-rays (3,107 with pneumothorax) were visually annotated by radiologists. After training, researchers tested the performance of the algorithms on a similar collection of labeled X-rays that had never been seen by the algorithms and analyzed their success at detecting images showing pneumothorax, without any human guidance.

Researchers found their algorithms were able to detect the majority (80%) of images showing a moderate or large pneumothorax, while correctly categorizing 97% of images without pneumothorax as “negative.”

Overall, these findings show that when given enough high-quality training data, computer algorithms are capable of detecting pneumothorax on a chest X-ray with sufficient accuracy to help prioritize images for rapid review by physicians. This means that algorithms like these could potentially be used by radiologists as a tool to increase the speed with which a serious pneumothorax is detected. Rapid detection and communication with treating physicians may result in faster treatment of pneumothorax, potentially reducing the harm of a serious medical problem.

This study was recently published in November 2018 in PLOS Medicine. Andrew Taylor, MD, PhD, associate professor of Clinical Radiology was lead author on the study along with Vice Chair for Informatics John Mongan, MD, PhD as senior author. Second author Clinton Mielke, PhD of CDHI was the data scientist on the project.

Related People