Point-of-Care AI for Chest X-Rays Accurately Classifies Optimal and Suboptimal Chest Radiographs

Chest radiographs (CXR) are the most common imaging test, accounting for nearly 40% of all imaging examinations. This popularity is due to their accessibility, practicality, low cost, and moderate sensitivity in diagnosing pulmonary, mediastinal, and cardiac issues. However, there is significant variability in CXR interpretation among radiologists. Higher quality images could lead to more consistent and reliable readings, but suboptimal CXRs can hinder the detection of critical findings. Now, radiologist-trained artificial intelligence (AI) models can accurately classify optimal and suboptimal CXRs, potentially enabling radiographers to repeat poor-quality scans when necessary.

Radiologists at the Massachusetts General Hospital and Harvard Medical School (Boston, MA, USA) have developed AI models that can distinguish between optimal and suboptimal CXRs and provide feedback on the reasons for suboptimality. This feedback, offered at the front end of radiographic equipment, could prompt immediate repeat acquisitions when needed. The radiologists utilized an AI tool-building platform to create their model that allows clinicians to develop AI models without prior expertise in data sciences or computer programming. This software could help reduce variability among radiologists.

The development of the model involved 3,278 CXRs from five different sites. A chest radiologist assessed the images and identified the reasons for their suboptimality. These anonymized images were then uploaded to an AI server application for training and testing. The model's performance was evaluated based on its area under the curve (AUC) for distinguishing between optimal and suboptimal images. Reasons for suboptimality included missing anatomy, obscured thoracic anatomy, inadequate exposure, low lung volume, or patient rotation. The AUCs for accuracy in each category ranged from .87 to .94.

The model demonstrated a consistent performance across age groups, sexes, and various radiographic projections. Importantly, the categorization of suboptimality is not time-consuming and it takes less than a second per radiograph per category to classify an image as optimal or suboptimal, according to the experts. The team has suggested that this could speed up the repeat process as well as streamline manual audits, which are typically laborious and time-consuming.

“An automated process using the trained AI models can help track such information in near time and provide targeted, large-scale feedback to the technologists and the department on specific suboptimal causes,” the group explained, adding that in the long-term this feedback could reduce repeat rates, saving time, money and unnecessary radiation exposures.

Related Links:
Mass General

Visit expo >

Ultrasound Table

Women’s Ultrasound EA Table

Radiation Therapy Treatment Software Application

Elekta ONE

Multi-Use Ultrasound Table

Clinton

NMUS & MSK Ultrasound

InVisus Pro