Development and evaluation of a deep learning-assisted diagnostic support system for radiographer preliminary clinical evaluation of intracranial hemorrhage

Background Intracranial hemorrhage is life-threatening and requires prompt and accurate diagnosis. Non-contrast head computed tomography is the standard first-line examination, but detecting small hemorrhages and classifying multiple subtypes require substantial expertise. Workforce shortages and increasing diagnostic workloads, especially in emergency settings, further challenge timely decision-making. Artificial intelligence (AI)-assisted interpretation has shown promise for improving accuracy and efficiency. This retrospective study evaluated the effect of AI assistance on the diagnostic performance of radiologic technologists (RTs). Methods We analyzed the data for 100 non-contrast head computed tomography examinations (50 positive and 50 negative for hemorrhage) obtained from the Japan Medical Image Database. The interpretations of the five RTs (5–12 years of experience) with and without AI assistance were compared with those of two radiologists. The detection targets were intraparenchymal, intraventricular, subarachnoid, subdural, epidural, and any hemorrhages. We calculated the Area Under the Receiver Operating Characteristic Curve (AUC), accuracy, sensitivity, and specificity. The differences in the AUC for the AI-assisted and unassisted readings were tested using the DeLong method with Bonferroni correction. Results Significant AUC improvements were observed for five of the 30 reader–task comparisons (17%) after Bonferroni correction. These improvements were all related to intraventricular ( p = 0.0001 to 0.0071) and subdural ( p = 0.0022 to 0.0071) hemorrhages. Conclusion AI assistance significantly improved RT detection of challenging subtypes such as intraventricular and subdural hemorrhages. However, it did not improve the diagnostic accuracy for detecting any hemorrhage overall ( p = 0.0689 to 0.9669). AI can strengthen the role of RTs within task-sharing models and help stabilize preliminary assessments, especially in emergency care and resource-constrained environments.