[Value of artificial intelligence in the improvement of diagnostic consistency of radiology residents].

OBJECTIVE

To explore the value of artificial intelligence (AI) in improving the detection rate of traumatic rib fractures by radiologist residents and the consistency among different readers.

METHODS

Chest CT images of 393 patients with acute trauma from China-Japan Union Hospital of Jilin University (hospital 02) and Shanghai Ninth People' s Hospital (hospital 03) were collected in this research. The consensus achieved by three radiology experts was regarded as the reference standard. All the images assigned to three hospitals: Peking University First Hospital (hospital 01), hospital 02 and hospital 03, and were then randomly divided into two groups (group A and group B: group A included 197 patients, and group B included 196 patients). Each group was read by one radiologist resident from each hospital for rib fracture detection. Each case was read twice by the same radiologist, with and without the assistance of the AI ["radiologist-only" reading and "radiologist + AI" reading]. The detection rates of different types of rib fractures (displaced fractures and occult fractures) were compared between "radiologist-only" reading and "radiologist + AI" reading. The consistencies of different radiologists with different reading methods were evaluated.

RESULTS

The detection rates of displaced rib fractures and occult rib fractures by "radiologist + AI" reading were significantly higher than those read by "radiologist-only" reading (94.56% vs. 78.40%, 76.60% vs. 49.42%, P < 0.001). For "radiologist-only reading", the Kappa coefficients of the radiologists between hospital 01 and hospital 03 were slightly greater than 0.4 (indicating moderate consistency), the coefficients of the radiologists between hospital 01/hospital 02 and hospital 02/hospital 03 were less than 0.4 (indicating poor consistency). The Phi coefficients of the radiologists among different hospitals were all less than 0.6 (indicating moderate correlation). With "radiologist + AI" reading, the Kappa and Phi coefficient among the radiologists in dif-ferent hospitals were greater than or equal to 0.6 (indicating good consistency and correlation).

CONCLUSION

AI software can be used to automatically detect suspected rib fracture lesions, which helps to improve the detection rate of fracture lesions and the consistency among different readers.