AI Case Study

King's College London researchers automate real-time triage of X-ray radiograph reports with high accuracy

Researchers from King's College London devise a natural language processing (NLP) system which can extract text from handwritten radiograph reports. This is intended to automatically extract and triage radiographs in real-time to address backlog issues.

Industry

Healthcare

Healthcare Providers And Services

Project Overview

"Thee system operates in real time: When a radiograph is acquired, it is processed by the deep CNN and assigned a predicted priority level. It is then inserted into a “dynamic re- porting queue” on the basis of its predicted urgency and the waiting time of other already queued radiographs." The NLP system extracted text which was then classified with 15 radiological labels based on a rules-system. "These labels reflected the most common and clinically important radiographic findings within our data set. These were then mapped onto four clinical prioritization levels selected to reflect our current reporting practice, as follows: (a) critical, requiring an immediate report due to a clinically critical finding (eg, pneumothorax); (b) urgent, requiring a report within 48 hours due to a clinically important but not critical finding (eg, consolidation); (c) non-urgent, requiring a report within the standard departmental turnaround time due to nonclinically important findings (eg, hiatus hernia); and (d) normal (ie, no abnormalities on radiograph)."

Reported Results

"The NLP system was able to extract the presence or absence of almost all the radiologic findings within the free-text reports with a high degree of accuracy. Previously published studies have investigated the potential of NLP and computer vision techniques in the classification of radiographs (6,19,20) but not for real-time prioritization, which was our primary aim."

Technology

"Annotation of the radiographs was automated by developing an NLP system that was able to process and map the language used in each radiology report (16). The architecture of our NLP system was somewhat similar to the architecture described by Cornegruta et al (17) and Pesce et al (18). The computer vision system was implemented on the basis of ordinal regression models, making use of two deep CNNs for the automatic extraction of imaging patterns directly from pixel values. All 329698 images in the training set were used for end-to-end training of the convolutional networks".

Function

R And D

Core Research And Development

Background

"Ever increasing clinical demands on radiology departments worldwide have challenged current service delivery models, particularly in publicly funded health care systems. In some settings, it may not be feasible to report all acquired radiographs in a timely manner, leading to large backlogs of unreported studies. For example, the United Kingdom estimates that, at any time, 330,000 patients are waiting more than 30 days for their reports. Therefore, alternative models of care should be explored, particularly for chest radiographs, which account for 40% of all diagnostic images worldwide".

Benefits

Data

"470 388 fully anonymized institutional adult chest radiographs acquired from 2007 to 2017. The 413403 consecutive radiographs acquired before April 1, 2016, were separated into training (n = 329 698, 79.7%), testing (n = 41 407, 10%), and internal validation (n = 42 298, 10.2%) sets, ensuring that the distribution of age and abnormalities within each subset matched the entire data set."