Cervical spine fractures can lead to severe complications if left undetected, making accurate diagnosis a critical component of trauma care. However, interpreting CT scans in high-pressure environments often results in missed fractures, especially when scans involve thousands of images. Traditional re-evaluation methods using human reviewers fail to capture all diagnostic errors, particularly subtle fractures. This gap has driven the development of machine learning tools capable of identifying fractures overlooked by radiologists. By integrating advanced algorithms, ML offers an opportunity to enhance diagnostic accuracy, highlight common blind spots and improve patient outcomes.
Advances in AI for Medical Imaging
Machine learning has emerged as a transformative tool in medical imaging, enabling improved diagnostic precision and enhanced efficiency in radiological workflows. In the context of cervical spine trauma, the RSNA 2022 AI Challenge highlighted the capabilities of ML models in identifying fractures missed by radiologists. These models leverage large datasets and advanced algorithms to analyse CT scans systematically, with a two-stage process comprising vertebral segmentation and fracture classification. By focusing on regions of interest and generating heat maps, these models direct radiologists’ attention to subtle findings that might otherwise be overlooked.
A recent study investigated the use of seven award-winning models to detect fractures in 6,378 CT scans that were initially interpreted as negative for cervical spine fractures. The ML models identified 356 scans with probable fractures, which were reviewed by neuroradiologists to confirm the presence of true missed fractures. This process revealed 40 examinations involving fractures at specific vertebral sites, particularly the C7 transverse process, C5 spinous process and C6 spinous process. These locations, often at the periphery of imaging series, are known to be prone to human perceptual errors.
One of ML’s key advantages lies in its ability to consistently analyse vast numbers of images without fatigue. Radiologists interpreting trauma scans often face high workloads and tight turnaround times, factors that contribute to missed diagnoses. The results of this study demonstrate how ML can augment human expertise, serving as a concurrent reader to flag potential fractures, particularly in emergency settings. This combination of human and AI collaboration could significantly reduce diagnostic errors while improving workflow efficiency.
Clinical Significance of Missed Fractures
Although the overall rate of missed fractures in the study was relatively low (0.6%), their potential clinical significance underscores the importance of accurate detection. Of the 40 scans with ML-detected missed fractures, 15 were deemed clinically significant. These cases required interventions such as MRI, surgical consultation or immediate cervical collar immobilisation. Notably, fractures involving the transverse and spinous processes accounted for a substantial proportion of missed diagnoses. While such fractures may not always appear critical, they can lead to serious complications if left untreated, including ligamentous injuries, vascular damage or nerve root compression.
The findings reveal that ML not only identifies missed fractures but also helps prioritise those that warrant further clinical attention. For instance, transverse process fractures extending into the foramen transversarium were classified as clinically significant in several cases, requiring additional imaging to assess associated vascular injuries. By highlighting these fractures, ML models contribute to earlier and more accurate diagnoses, ensuring appropriate and timely treatment to prevent further complications.
The integration of ML into radiological workflows has the potential to reduce satisfaction-of-search errors, where radiologists overlook additional findings after detecting an initial abnormality. In this study, many missed fractures occurred in scans obtained as part of whole-body trauma protocols, which involved reviewing thousands of images. Such demanding cases are particularly prone to diagnostic oversights. ML tools can help mitigate this issue by drawing attention to areas of concern, thereby improving overall diagnostic accuracy and reducing the likelihood of errors.
Integrating Machine Learning into Radiology Workflow
To maximise the benefits of ML in clinical practice, careful integration into existing radiology workflows is essential. ML algorithms can act as triage tools, flagging potential fractures and directing radiologists’ focus to regions at higher risk of being missed. However, reducing false-positive rates remains critical to prevent alarm fatigue, which could undermine the utility of these systems. Enhancing the specificity of ML tools will be an important step towards their widespread adoption.
Additionally, structured reporting systems and checklists can complement ML-driven solutions by encouraging thorough examination of CT images. By incorporating reminders to focus on areas prone to missed fractures—such as transverse and spinous processes—radiologists can improve their search patterns. Checklists have been shown to enhance radiologists’ thoroughness, particularly in high-stress scenarios like emergency trauma care. Combining these strategies with ML assistance offers a robust approach to minimising diagnostic errors.
Future workflows may involve hybrid models that combine the strengths of radiologists and ML systems. While ML excels at detecting subtle patterns across large datasets, radiologists bring clinical expertise and contextual understanding to refine diagnoses. This collaborative approach ensures both efficiency and accuracy, meeting the increasing demands of modern healthcare systems.
Machine learning offers a powerful solution to the persistent challenge of missed cervical spine fractures in CT imaging. By systematically identifying fractures overlooked by radiologists and assessing their clinical relevance, ML models provide valuable insights into common diagnostic errors and their potential impact on patient care. The study highlights the importance of integrating ML tools into radiology workflows to enhance diagnostic accuracy, reduce perceptual errors and ensure timely treatment for trauma patients. With further refinement and careful implementation, ML has the potential to revolutionise radiological practice, offering a future where no critical fracture goes undetected.
Source: American Journal of Radiology
Image Credit: iStock