This diagnostic study evaluated the performance of an AI algorithm vs. radiology residents in interpreting anteroposterior (AP) frontal chest radiographs in emergency and inpatient settings. The AI model, trained on 342,126 frontal chest X-rays from multiple hospitals, was tested against five U.S.-based third-year radiology residents using a separate dataset of 1,998 X-rays with ground truth labels established via triple consensus adjudication. The AI achieved a mean AUC of 0.807 (training) and 0.772 (field testing), with PPV (0.730) and specificity (0.980) significantly exceeding those of radiology residents (PPV: 0.682, specificity: 0.973; P < .001), while sensitivity remained comparable (AI: 0.716, Residents: 0.720; P = .66). These findings suggest that AI algorithms can match or outperform radiology residents in preliminary CXR reads, particularly in routine cases, but expert oversight is still required for complex findings. Integrating AI into radiology workflows could enhance efficiency, improve diagnostic accuracy, and address resource shortages in emergency care.
Published 23 Jul 2020
Comparison of Chest Radiograph Interpretations by Artificial Intelligence vs Radiologists in the diagnosis of pulmonary TB
Author: Joy T. Wu, MBChB, MPH;
Ken C. L. Wong, PhD;
Yaniv Gur, PhD;
Nadeem Ansari, MS;
Alexandros Karargyris, PhD;
Arjun Sharma, MD;
Michael Morris, MD;
Babak Saboury, MD;
Hassan Ahmad, MD;
Orest Boyko, MD, PhD;
Ali Syed, MD;
Ashutosh Jadhav, PhD;
Hongzhi Wang, PhD;
Anup Pillai, PhD;
Satyananda Kashyap, PhD;
Mehdi Moradi, PhD;
Tanveer Syeda-Mahmood, PhD
Authors
Joy T. Wu, MBChB, MPH; Ken C. L. Wong, PhD; Yaniv Gur, PhD; Nadeem Ansari, MS; Alexandros Karargyris, PhD; Arjun Sharma, MD; Michael Morris, MD; Babak Saboury, MD; Hassan Ahmad, MD; Orest Boyko, MD, PhD; Ali Syed, MD; Ashutosh Jadhav, PhD; Hongzhi Wang, PhD; Anup Pillai, PhD; Satyananda Kashyap, PhD; Mehdi Moradi, PhD; Tanveer Syeda-Mahmood, PhD
