JMIR Publications

ABSTRACT

Background:

Traumatic brain injury (TBI) is one of the leading causes of disability and mortality worldwide. Timely diagnosis is essential to minimize secondary injuries and improve patient outcomes. However, the manual evaluation of TBI computed tomography (CT) scans can result in delays of diagnosing severe cases. This issue is compounded in smaller hospitals where limited radiological expertise leads to general practitioners interpreting imaging data. Recent advancements in artificial intelligence (AI) have permitted automated support systems for accurate triaging.

Objective:

This study presents a comprehensive review of AI techniques applied to CT scans for TBI assessment.

Methods:

This study followed the guidelines specified in the Preferred Reporting Items for Systematic reviews and Meta-Analyses (PRISMA) 2020

Results:

By screening the 674 publications found, we identified 101 studies for evaluation. We grouped these 101 studies into four categories: Binary TBI Classification (TBI / Non-TBI), Hemorrhage Detection (Classification, Localization, Segmentation, Quantization), Midline Shift Measurement, and Increased Intracranial Pressure Estimation.

Conclusions:

We have highlighted the progress made in the application of AI for TBI assessment on CT scans. It is evident that there is a focus on hemorrhage detection accounting for 85% of the total studies (86 of the 101). Despite advances in research to improve the performance of AI models in TBI assessment, several key challenges must be addressed to improve deploying these models into clinical workflows in real-world settings. These challenges include the need for more inclusive models that address clinical TBI conditions. In addition, the development of large-scale, publicly available datasets to mitigate overfitting and ensure generalization. Most importantly, a gap in the literature is the annotation of data. To improve the development of these models, we need to explore alternative learning strategies to efficiently reduce annotation burdens. Given clinicians have the expertise needed for accurate annotation, their clinical commitment leaves minimal time for labelling data. Furthermore, ensuring transparency and interpretability of AI models is essential for successful integration into clinical practice and clinical acceptance. Finally, clinical validation with healthcare professionals is critical to confirm the practical utility and reliability of AI-based TBI diagnostic tools. By addressing these challenges, AI can become a more effective tool in TBI assessment.