DEEP LEARNING IN RADIOGRAPHIC TRIAGE: WORKFLOW OPTIMIZATION TO ADDRESS THE RADIOLOGIST WORKFORCE CRISIS

Background: The global radiologist workforce faces a systemic crisis where imaging volume growth significantly outpaces specialist capacity, reducing per-image interpretation time from 16.0 to 2.9 seconds. This chronic overload contributes to burnout rates between 34% and 39% and increases the risk of diagnostic errors when daily productivity is exceeded by approximately 21%. Methods: A comprehensive literature review examined peer-reviewed studies published between 2015 and 2025. The analysis focused on the efficacy and sociotechnical impact of deep learning (DL) models across four critical pathologies: intracranial hemorrhage (ICH), large vessel occlusion (LVO) stroke, pulmonary embolism (PE), and pneumothorax. Results: DL models, primarily Convolutional Neural Networks and Vision Transformers, demonstrate high diagnostic accuracy, with pooled sensitivities and specificities frequently reaching 90%. "Active reprioritization" significantly reduces report turnaround times, yielding median savings of 12.3 minutes for PE and 20.5 minutes for stroke. For outpatient ICH, time-to-diagnosis dropped from 512 minutes to 19 minutes. In acute stroke care, AI facilitation resulted in a 30.2-minute reduction in door-to-treatment times and improved discharge NIHSS scores. Conclusions: DL triage serves as a vital sociotechnical intervention to preserve patient safety amidst diagnostic overload. Its primary clinical value resides in workflow orchestration rather than standalone diagnosis. Successful implementation requires integrated "human-in-the-loop" systems to mitigate automation bias and the cognitive time penalty associated with false positives.