Methods:

An AI model aimed at identifying TL fractures on plain radiographs was trained and tested on 567 single vertebrae images. Three expert spine surgeons established the Ground Truth (GT) using CT and MRI to confirm the presence of the fracture. From the test set, 12 cases (6 with a GT of fracture and 6 with a GT of no fracture, associated with varying levels of algorithm confidence) were selected and the corresponding SMs were generated and shown to 7 independent evaluators with different grade of experience; the evaluators were requested to: (1) identify the presence or absence of a fracture before and after the saliency map was shown; (2) grade, with a score from 1 (low) to 6 (high) the pertinency (correlation between the map and the human diagnosis), and the utility (the perceived utility in confirming or not the initial diagnosis) of the SM. Furthermore, the usefulness of the SM was evaluated through the rate of correct change in diagnosis after the maps had been shown. Finally, the obtained scores were correlated with the algorithm confidence for the specific case