Analysis of Rib Fractures Detected by AI-Initial Experience

Purpose: The purpose of our study was to evaluate the accuracy of AI software that automatically identifies rib fractures.

Patients and Methods: Thirty CT exams marked as positive by the AIDOC software were review by two radiologists. Each positive location was evaluated as either true positive (TP) or false positive (FP). True positive fractures were graded as acute, subacute, old and pathologic. Each exam was evaluated for motion artifacts and the existence of unidentified fractures. The indication for each exam was registered as either trauma of other indications.

Results: 4318 cases were evaluated by the software out of which 146 (3.3%) were prioritized with suspected rib fractures. Forty-three percent of patients were scanned due to trauma. There was an average of 2.85 markings per patient. FP were seen in 3.5% of cases. Sixty-three percent of fractures were acute, 9% subacute, 16% were old and 10% pathologic. Motion artifacts were noted in 20% and in half of the FP locations. In 43% of prioritized cases we found other unidentified fractures.

Conclusion: AI based software that automatically prioritizes rib fractures is reliable in identifying fractures of varying ages and types. There is still need to evaluate the cases that were not prioritized to evaluate the existence of non-detected fractures.