Evaluation of an artificial intelligence-based software device for detection of intracranial haemorrhage in teleradiology practice

This retrospective diagnostic accuracy study aimed to assess the performance of an AI-based software device in detecting intracranial haemorrhage (ICH) on non-contrast CT head (NCCTH) scans within a large teleradiology practice. The rationale behind the study was to evaluate the potential of AI algorithms in supporting radiologists by improving diagnostic accuracy and efficiency for urgent cases. A randomly selected sample of 1,315 NCCTH scans from adult patients (aged ≥18 years) referred from 44 hospitals across the UK over four months was retrospectively evaluated by 30 auditing radiologists alongside the AI system. Among these scans, 112 (8.5%) were confirmed to have ICH. The results showed a high level of agreement between the AI system and the radiologists, with an overall agreement of 93.5%, a positive percent agreement of 85.7%, a negative percent agreement of 94.3%, and a Gwet’s AC1 score of 0.92, indicating very good reliability. Most false negatives were due to subtle subarachnoid haemorrhages, while false positives were primarily caused by motion artefacts. In conclusion, the study demonstrated that the AI system performed comparably to radiologists in detecting ICH, highlighting its potential as a supportive tool in teleradiology for enhancing diagnostic accuracy and efficiency.