Traumatic brain injury (TBI), often referred to as the “silent epidemic” [1-3], affects an estimated 64‐74 million people globally each year [3]. In total, 176 American people die each day due to causes related to TBI as per the United States Centers for Disease Control and Prevention report “Multiple Cause of Death Data” released in 2020 [4]. TBI pathophysiology is generally categorized into 2 stages: primary and secondary injuries. Primary injuries are irreversible injuries that happen after the first physical trauma to the brain [5,6]. Secondary injuries develop over time due to complex molecular, metabolic, and inflammatory responses. These responses can lead to elevated intracranial pressure (ICP), cerebral vasospasm, ischemia, herniation, infarction, and midline shift (MLS) [7-9]. Unlike primary injuries, secondary injuries present a critical window of opportunity for intervention, referred to as the “golden hours.” Timely assessment and medical treatment during the golden hours following the trauma can significantly reduce or delay the progression and severity of these injuries [10,11]. Therefore, minimizing the time to treatment is essential in improving the outcomes of patients with TBI.

In recognition of this urgency, clinical management of suspected TBI begins immediately upon the patient’s arrival at the hospital. Figure 1 illustrates this clinical workflow. When a patient arrives at the hospital with suspected TBI, the emergency team first stabilizes the patient by performing the primary survey, which is designed to assess and treat life-threatening injuries in a timely fashion, using the systematic Airway, Breathing, Circulation, Disability, Exposure algorithm. The algorithms include an assessment of the patient’s neurological status, performed by evaluating consciousness using a Glasgow Coma Scale (GCS) and assessing the function of the central and peripheral nervous systems through motor strength, sensory function, reflexes, pupils’ examination, etc, which may provide signs of TBI. Based on these assessments, the emergency team determines the patient’s priority for care. Further imaging, including a brain scan as part of whole-body imaging, is requested when there are signs of TBI, such as low GCS scores, weaker motor strength, or unequal or unreactive pupils. Radiologists review images to confirm the diagnosis and provide detailed reports, enabling subspecialized teams to manage serious injuries.

Several imaging diagnostic tools are available for diagnosing TBI, depending on the care available, including computed tomography (CT) scans, magnetic resonance imaging (MRI) scans, positron emission tomography (PET) scans, ultrasound, and transcranial Doppler. These alternative tools have their own advantages and disadvantages. A comparison of these tools across the dimensions relevant to hospital settings is presented in Multimedia Appendix 1. It includes an evaluation on noninvasive, rapid results, cost, high-resolution anatomical imaging, accessibility, availability for emergency use, contraindications, and clinical utility. While MRI provides high sensitivity for detecting subtle injuries such as diffuse axonal injury or microhemorrhages, its use is often limited in acute settings due to longer scan times, higher costs, and reduced availability. Ultrasound and transcranial Doppler are noninvasive and useful for evaluating blood flow dynamics in brain vessels; however, they lack the spatial resolution to detect parenchymal injuries or structural abnormalities within the brain. Additionally, the accuracy of ultrasound or Doppler imaging can be subjective, as it depends on uncontrolled external factors, such as the technician’s expertise, patient movement, and the scanning technique used. Consequently, CT is the preferred first-line imaging modality for patients with trauma to diagnose TBI, as it is noninvasive, provides quick results, is widely available, and has few contraindications [12-14].

Despite its advantages, the reliance on CT imaging in TBI assessment presents its own set of challenges. First, in current practices, clinicians (in this paper, the term “clinicians” refers broadly to health care professionals involved in the TBI assessment, including general practitioners, radiologists, neurosurgeons, and other relevant specialists) typically perform assessments manually based on individual CT slices. Radiologists are inundated with medical images daily, with an estimated interpretation of hundreds of scans per shift [15]. In a busy hospital, this translates to thousands of scans requiring detailed analysis [16], creating an overwhelmingly heavy workload. Moreover, manual methods often fail to accurately and comprehensively quantify critical injury characteristics, such as hematoma size and volume, the degree of MLS, the extent of brain contusions, and indicators of diffuse axonal injury. Studies show that the average rate of diagnostic errors in radiology reports ranges from 3% to 5%, which translates to roughly 40 million imaging diagnostic errors each year [17]. Exacerbating this problem is an uneven distribution of skilled health care professionals and resource limitations, resulting in notable differences in care delivery and decision-making processes across various health care settings. For example, the Alfred Trauma Center in Australia has neurosurgeons available to review CT scans in real time via a dedicated viewing system. This enables the detection of severe TBI indicators in real time, allowing patients to be immediately transferred for surgical intervention. In contrast, smaller hospitals often face resource constraints, such as the unavailability of on-site radiologists. In such cases, nonspecialized or junior staff must review and assess the CT scans to identify severe cases that require immediate attention or necessitate the assistance of tertiary centers and/or escalation to a remote radiologist. This responsibility is particularly challenging when (1) other medical conditions mimic TBI symptoms or (2) the nonspecialized medical staff have limited experience with complex neuroimaging cases. Such scenarios are prone to inter- and intraobserver variability, leading to inconsistent interpretations and reduced diagnostic accuracy [17,18].

These challenges highlight the urgent need for automated, consistent solutions to support neuroimaging assessments, thereby alerting to concerning images, reducing interpretation time, and ensuring consistency in TBI assessments. Computerized tools could be transformative for smaller, rural hospitals to support health care workers when specialized neurosurgical and neurological care resources are scarce [19,20]. To address these issues, researchers are leveraging AI, particularly computer vision techniques, to automate the process of analyzing CT scans, thereby supporting medical staff in reducing the time required to identify severe TBI cases. AI-driven models can accurately detect and quantify critical findings, such as intracranial hemorrhages (ICHs) and MLSs [21]. In addition, these systems can support general practitioners and radiologists by reducing the time to identify and characterize TBIs in a scan from 15‐30 minutes per case [22] to seconds [23]. These advancements underscore the potential of integrating AI into the TBI assessment workflows in trauma units.

However, to develop an effective AI tool that supports clinicians in TBI assessment, the proposed AI framework must be tailored to address the challenges in integrating AI into clinical settings. This study aims to summarize existing AI approaches for detecting and quantifying CT-based abnormalities in TBI, highlighting their capabilities and limitations. Our research questions are as follows:

We adopted the PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) guidelines for a rigorous approach in selecting literature. However, this study constitutes a systematized review rather than a full systematic review, and certain limitations in our coverage are acknowledged. A screening process excluded studies due to availability, accessibility, and relevance to this review. While we have strived for a comprehensive and unbiased selection, the process may still reflect criteria that introduce potential biases. Furthermore, the present analysis does not evaluate or compare the performance of AI models against real-world clinical gold standards, as such validation falls beyond the scope of this review. Consequently, bias assessment frameworks such as QUADAS-2 (Quality Assessment of Diagnostic Accuracy Studies-2) and PROBAST (Prediction Model Risk of Bias Assessment Tool) are also not applied.

This paper is structured as follows: the Methods section outlines the required background knowledge and presents our methodology. The Results section presents the review results of AI applications in TBI assessments, including datasets and data processing, TBI detection, hematoma localization, MLS measurement, and ICP estimation. Finally, we discuss current limitations and propose future research directions to advance AI integration into clinical practice in the Discussion section.

The research questions are addressed as follows: RQ1 is addressed in the Search Results section, RQ2-RQ3 are addressed in the TBI Datasets and Data Preprocessing Pipeline section, RQ4 is addressed in the Binary TBI Classification through ICP Estimation sections, RQ5 is synthesized across the Results and Discussion sections, and RQ6 is discussed explicitly in the Discussion section.

TBI is a description of structural abnormalities (damage) in the brain such as hematomas, MLSs, and increased ICP. Hematomas are one of the most common complications in TBI and are broadly classified into extra-axial and intra-axial types based on their location relative to the brain tissue. Extra-axial hematomas include epidural hematomas (EDHs), subdural hematomas (SDHs), and subarachnoid hemorrhages (SAH), while intra-axial hematomas comprise intraparenchymal hemorrhages (IPHs) and intraventricular hemorrhages (IVHs) [24,25]. CT remains the primary imaging modality for detecting and characterizing these hemorrhagic lesions. Clinicians assessing CT scans with presented hematomas have a visualization of the size, location, and extent of bleeding. This information serves as vital prognostic indicators [26,27] determining the urgency and type of intervention, such as surgical evacuation or conservative management.

MLS is another critical quantitative measure in TBI assessment [26,28]. It arises when elevated ICP or mass effects from conditions such as hematomas or brain swelling force brain structures away from their normal anatomical positions. MLS is assessed in CT scans by measuring the displacement of midline anatomical landmarks, such as the septum pellucidum or falx cerebri [29]. Clinicians measure the MLS by identifying the ideal midline (iML), which is often aligned with the falx cerebri’s attachment to the skull, then calculating the distance between the iML and a specific brain structure that has shifted, such as the septum pellucidum.

ICP is another determinant of TBI severity. Elevated ICP is observed in brain injuries due to the mass effect from ICH [28]. An increase in ICP decreases brain blood flow pressure (ie, cerebral perfusion pressure), resulting in compression, deformation, and herniation of brain tissue, which can exacerbate the injury [30,31]. Clinical intervention is necessary when ICP exceeds 22 mm Hg; levels above this threshold lead to an increased risk of fatality [32]. Clinical guidelines recommend ICP monitoring for patients with a GCS score <8 and an abnormal head CT. Abnormal CT findings include mass lesions (eg, hematoma or contusion), cerebral edema or swelling, MLS, and compressed basal cisterns, particularly the perimesencephalic cisterns [33]. Currently, ICP measurements are invasive, whereby clinicians adopt ventricular or parenchymal pressure monitoring [34,35]. Despite this direct approach in measuring ICP, CT imaging allows for indirect ICP assessments. Radiological indicators for assessing ICP through CT scans include compressed ventricles, effaced sulci, and MLS, all suggestive of elevated ICP.

AI has shown promise in streamlining TBI characteristics from CT scans by adopting computer vision techniques for detecting brain structural abnormalities. This can support clinicians in triaging TBI cases, reducing the time to assessment and intervention for severe cases [23]. In addition, AI can support clinicians in reducing diagnostic errors [36]. One common AI technique is classification, where an AI model is trained to label a CT scan with a predefined category (eg, normal or abnormal), based on examples it has seen before. Another helpful technique is object detection, which identifies the presence of an abnormality such as a hematoma and provides additional information such as its location in the scan. Segmentation is another AI technique that separates the image into different regions, clearly outlining the edges of abnormal areas. This makes it easier to study the shape and structure of parts of the brain such as lesions. Building on segmentation, quantification techniques measure specific details of these abnormalities, such as the volume of a hematoma or the distance of the brain’s MLS.

This study was conducted as a systematized review following the PRISMA 2020 reporting guideline [37]. While the literature search, study selection, and reporting were performed using systematic and reproducible procedures, this review does not meet all methodological criteria required for a full systematic review. PRISMA guidance was therefore used as a framework to promote transparency and completeness of reporting, rather than to imply full compliance with all requirements of a registered systematic review (Checklist 1). Specifically, no formal protocol was registered in a public database such as PROSPERO prior to conducting the review. Study screening and data extraction were performed by a single reviewer (VTNP). Furthermore, the review focuses on mapping AI tasks, methodologies, datasets, and evaluation practices rather than estimating overall diagnostic performance of those models. The present analysis does not evaluate or compare the performance of AI models against real-world clinical gold standards, as such validation falls beyond the scope of this review. Consequently, bias assessment frameworks such as QUADAS-2 and PROBAST are also not applied. To address potential biases relevant to AI research, we focused on examining dataset characteristics, annotation transparency, and validation strategy. The implications of these factors for interpretation and generalizability are discussed explicitly in the Discussion section.

The search was conducted across 4 major scientific databases including PubMed, Scopus, Web of Science, and IEEE Xplore, covering publications available up to December 31, 2025. Search queries combined keywords related to AI, machine learning (ML), deep learning (DL), CT imaging, and TBI, supplemented with MeSH terms. A detailed list of the search queries used is reported in Table S1 in Multimedia Appendix 2.

The screening strategy for this review was designed to select studies aligned with the research objectives; refer to Table S2 in Multimedia Appendix 2 for our inclusion and exclusion criteria. This review focused exclusively on AI approaches applied to noncontrast head CT imaging. Studies incorporating multimodal inputs such as clinical scores, physiological signals, or laboratory data were excluded to maintain a focused analysis of imaging-specific methodologies and evaluation practices. This restriction was intended to limit methodological heterogeneity and to reflect the central role of CT imaging in acute TBI assessment workflows.

Eligible studies included studies that (1) were published as a peer-reviewed journal paper or conference paper, (2) used CT scans as input data, and (3) involved the development or application of AI algorithms for TBI assessment, including tasks such as hemorrhage segmentation, MLS measurement, or ICP estimation.

Studies were excluded based on (1) duplication or (2) retracted publications; (3) imaging modalities used other than CT scans (eg, MRI and electroencephalogram); (4) nonhuman studies; (5) conditions unrelated to TBI (eg, stroke, brain tumors, or other neurological disorders); (6) infants or pediatric populations, as TBI and ICP differ from adults; (7) non-English language publications; (8) nonoriginal research (eg, reviews, meta-analyses, and secondary analyses) that are limited in research details; (9) statistical analysis, therapeutic techniques, and pathological investigations not in the field of AI; or (10) nonrelevant outcomes (eg, hematoma expansion, patient prognosis, and risk prediction). These exclusion criteria ensure that the review remains concentrated on AI applications for detecting TBI-related abnormalities.

All records retrieved from the database searches were exported to Zotero—a reference management tool, where duplicates were identified and removed. Titles and abstracts were screened against the eligibility criteria, followed by full-text screening of potentially relevant papers. Study selection was performed by one reviewer; no automation tools were used for decision-making, and no second reviewer screening was performed, which may increase the risk of selection bias. To manage the scope of this systematized review, an exceptionally stringent filtering approach was applied during the title and abstract screening phase. Studies exhibiting any ambiguity regarding imaging modality (eg, lack of explicit noncontrast CT confirmation), patient demographics, or the application of AI methodologies were excluded at this initial stage. Consequently, the 101 papers that advanced to the full-text assessment had already strictly satisfied all inclusion criteria based on their abstracts, resulting in zero subsequent exclusions during the full-text confirmation phase.

By screening the 674 publications found, we identified 101 studies for evaluation. To address RQ1, the reviewed studies are first grouped according to the primary TBI assessment task they aim to solve. We grouped these 101 studies into 4 categories: binary TBI classification (TBI or non-TBI), hemorrhage detection (classification, localization, segmentation, and quantization), MLS measurement, and increased ICP estimation. It is evident that the literature heavily focuses on hemorrhage detection, which accounts for 85% (86/101) of the total studies. The remaining studies are distributed across MLS measurement (8/101), binary TBI classification (TBI or non-TBI; 4/101), and increased ICP estimation (3/101). The remainder of this review focuses on AI applications in TBI, identifying the techniques applied and results presented in these 101 studies.

A structured data extraction form was designed a priori and piloted on a small subset of studies to ensure consistency and clarity. For each study included, one reviewer extracted the data. When information was unclear or missing, we relied on the published paper and supplementary materials; no contact with study authors was attempted.

From each study, we extracted:

When an outcome or methodological detail was ambiguously reported, we recorded the information as described by the authors and noted any uncertainties qualitatively in the synthesis. To accurately assess the reproducibility and completeness of the published literature, no contact with study authors was attempted to retrieve missing data or methodological justifications not included in the peer-reviewed texts. We treated the performance metrics reported by the primary studies as the main effect measures. Given the heterogeneity of tasks, datasets, and metrics, these effect measures were summarized descriptively rather than pooled quantitatively.

Because of substantial heterogeneity in clinical tasks (eg, binary TBI classification, hemorrhage detection and segmentation, MLS estimation, and ICP assessment), datasets (public vs institutional, sample sizes, and annotation strategies), and reported outcome metrics, we did not attempt a meta-analysis. Instead, we conducted a narrative synthesis. Studies were grouped by primary clinical task and, where relevant, by model family. Within each group, we compared model architectures, training strategies, preprocessing approaches, and reported performance metrics. To evaluate data preparation trends, we systematically extracted all reported preprocessing techniques from the included studies (eg, windowing, skull stripping, and normalization). These discrete techniques were then qualitatively grouped into conceptual categories and ordered temporally to synthesize a consensus, 9-stage sequential preprocessing pipeline, providing a conceptual framework that maps current developmental practices.

We did not formally evaluate publication bias or small-study effects, as the included studies were highly diverse and rarely reported sufficient detail for such assessments. Similarly, we did not calculate graded certainty of evidence (eg, using GRADE [Grading of Recommendations Assessment, Development and Evaluation]) because the primary aim was to describe the methodological landscape rather than to derive pooled effect estimates for a specific clinical outcome.

We did not conduct a formal risk-of-bias assessment using tools such as QUADAS-2 or PROBAST because our primary focus was to map and characterize AI methodologies, tasks, and evaluation practices rather than to synthesize pooled estimates of diagnostic accuracy or prognostic performance. Instead, we qualitatively appraised methodological aspects likely to affect bias and generalizability, including dataset size and diversity, transparency of inclusion and exclusion criteria, clarity of annotation procedures, internal versus external validation, and reporting of model development and evaluation.

By screening the 674 publications found, we identified 101 studies for evaluation. The overall selection process and study count at each stage are summarized in the PRISMA flow diagram (Figure 2).

Clinicians assessing TBI examine CT scans for TBI abnormalities. This has been translated into AI binary classification tasks that identify the presence (or absence) of TBI. Dusenberry et al [46] used an artificial neural network (ANN) model to classify CT scans by detecting acute CT findings indicative of TBI. Their model achieved an accuracy of 93.14%, a recall of 97.78%, and a negative predictive value of 98.08%. These results demonstrate the model’s effectiveness in minimizing missed diagnoses. However, the model’s precision of 88% suggests a high rate of false positives. Gan et al [47] proposed a more complex architecture that combined a ResNet backbone with a squeeze-and-excitation (SE) module and a long short-term memory (LSTM) network. The model was initialized with ImageNet-pretrained weights and fine-tuned for the TBI classification task. It achieved an accuracy of 95.9%, with a sensitivity of 93.3% and a specificity of 98.9%. Zhang et al [48] extended this approach by integrating a traditional VGG-S CNN with both an SE module and a pixel-wise correlation retaining (PCR) module. This architecture, termed VGG-SE-PCR, improved feature extraction and retention of spatial relationships in CT scans. Although the model achieved a slightly lower overall accuracy of 89.3%, it exhibited a high sensitivity of 98.6%, suggesting a strong capability to detect TBI. However, its relatively low specificity of 82.9% demonstrates vulnerability to false positives. Ellethy et al [49] introduced a multimodal framework that combined CT imaging with clinical data using a region-based CNN architecture. The model achieved a comparatively lower performance, with accuracy, sensitivity, and specificity values of 82.4%, 82%, and 82%, respectively. These results suggest that while clinical data can offer valuable context, their inclusion may also introduce noise or increase model complexity if not effectively integrated. Despite these findings, Ellethy et al [49] highlighted the importance of a holistic approach to TBI assessment, suggesting that combining multiple modalities could be key to capturing the multifaceted nature of TBI presentations.

Across binary TBI classification studies, performance differences are driven primarily by model complexity and input aggregation strategy rather than dataset size alone. First, models relying exclusively on CT imaging consistently outperform naive multimodal approaches that incorporate clinical variables without structured fusion mechanisms. This suggests that poorly aligned clinical features may introduce noise rather than complementary information. Second, early ANN-based approaches achieve high sensitivity and negative predictive value, reflecting suitability for rule-out triage in older cohorts, but rely on handcrafted features and lack spatial modeling capacity. Meanwhile, CNN-based architectures augmented with channel attention (SE blocks) and temporal modeling (LSTM) demonstrate the highest overall accuracy and specificity, indicating improved discrimination of subtle injury patterns across CT slices. However, most binary classification models are evaluated on internal datasets only, limiting conclusions about generalizability. Moreover, despite improvements in accuracy and sensitivity, the challenge of balancing specificity to reduce false positives remains an open research question. Table S1 in Multimedia Appendix 3 [46-49] summarizes the reviews in TBI binary classification.

Researchers have proposed various models to improve MLS estimation and can be classified into 2 categories: landmark-based approaches and DL-based approaches, where the former account for 25% of studies and the latter 75% (refer to Table S6 in Multimedia Appendix 3) [103-110].

Landmark-based models rely on anatomical cues and geometric relationships. Hooshmand et al [103] developed a system that first selects a number of slices among all the slices in a CT scan based on metadata, as well as information extracted from the images such as the number of separate objects counted in that slice and the area of skull in comparison to brain tissue. Then, the line that divides the slices into 2 equal right and left partitions is marked as iML. Finally, they performed ventricles segmentation to detect the actual midline (aML). The aML is defined as the line that passes between the left and right lateral ventricles or through the third ventricle. MLS is then calculated as the distance between the iML and aML. Their model achieved moderate performance: 75% sensitivity, 65% specificity, 53% precision, 68% accuracy, and an F₁-score of 0.62. However, this system focuses only on linear displacement and does not account for variations in tissue shape or volume. To address these limitations, Jiang et al [104] introduced a volumetric approach by proposing the midsurface shift (MSS). Figure 4 visualizes the difference between MLS and MSS estimations. MSS is defined as the ratio of midsurface volume to total brain volume, calculated from selected CT slices in a CT scan. First, they used 3D Chan-Vese segmentation combined with an active contour on contrast-adjusted CT scans to segment the brain mask. The midsurface is then manually annotated using biomarkers such as the anterior, the falx cerebri, the middle of the ventricle, and the posterior. The iML is defined by the intersections of the midsurface curve and the inner edge of the skull, and the volume of shift is calculated based on this iML. Finally, the ratio between the volume of shift and the brain volume is computed. While conceptually innovative, this study has not yet been quantitatively validated.

Contrarily, DL-based models learn discriminative features directly from imaging data. Wei et al [105] developed a regression-based line detection network that incorporates multiscale line detection, weighted line integration, and regression-based line refinement. Their method formulated the task of identifying the brain’s midline structure as a skeleton extraction problem and used a multiscale approach to capture both high-level semantic and low-level detailed information, along with a regression-based refinement. They achieved a mean line distance error of 1.17 (SD 0.72) mm and an F₁-score of 0.78 on the CQ500 dataset. However, testing on a private dataset resulted in a larger mean error of 4.15 (SD 3.97) mm and a lower F₁-score of 0.61, underscoring challenges in generalization across datasets. Yan et al [106] introduce a modified Keypoint R-CNN framework trained on a ResNet-FPN-50 backbone to detect 4 keypoints: (1) the anterior falx, (2) the posterior falx, (3) the anterior septum pellucidum near anterior commissure, and (4) the posterior septum pellucidum. The MLS is then calculated by the distance between the middle point of the connection of the anterior and posterior falx and the midpoint of the septum pellucidum. Their method reported a mean absolute error of 0.936 mm, sensitivity of 91.7%, specificity of 97.4%, and an intraclass correlation coefficient of 0.9899 compared to clinician assessments. This close agreement with expert evaluations positions their approach as highly competitive in clinical settings. Xia et al [107] designed a CNN framework with three stages: (1) U-Net for image alignment, (2) U-Net for midline segmentation, and (3) dynamic programming for pathfinding. The MLS is estimated by calculating the distance of midline shift (MLS-D) and the volume of midline shift (MLS-V). The horizontal distances between the aML and the iML are calculated on all shifted slices, with the largest distance chosen as the MLS-D. The MLS-V is calculated as the number of voxels between the aMLs and iMLs multiplied by the physical size of a voxel, whereby a voxel is defined as a discrete element of volume across the x, y, and z axes (x-spacing×y-spacing×z-spacing). They achieved a mean shift distance error of 1.14 (SD 0.91) mm and a mean shift area error of 0.88 (SD 0.79) cm², with AUC-ROC scores of 0.799 and 0.831 for MLS-D and MLS-V, respectively. Nag et al [108] also applied a U-Net–based approach to segment the hemispheres and define the aML based on boundary detection, estimating the iML by connecting the top-most and bottom-most points of the aML. They achieved a validation accuracy of 94.05%, with a pixel error of 1.294 mm, an area error of 66.4 mm², and a volume error of 253.73 mm³. Similarly, Wu et al [109] proposed a hemisphere-segmentation framework based on a 3D U-Net architecture. This network is divided into three components: (1) a hemisphere-segmentation network, (2) rectification learning, and (3) midline correction. They achieved an average surface distance of 0.84 (SD 0.4) mm and a mean Hausdorff distance of 9.1±7.5 mm on the CQ500 dataset.

Recently, several noninvasive methods for estimating ICP have been developed. However, the body of research in this area remains relatively limited, accounting for only about 2.9% of the reviewed literature. Pappu et al [112] proposed a semiautomatic technique that isolates brain tissue from cerebrospinal fluid (CSF) on CT scans. Each pixel is assigned an indicator value for CSF, blood, or brain tissue. The algorithm calculates the ratio of cerebrospinal fluid volume (CSFV) to the intracranial vault volume (ICVV). These computed ratios (CSFV/ICVV) were plotted against the recorded ICP values. Statistical analysis was performed using the Statistics Toolbox in MATLAB. They found that a CSFV/ICVV ratio exceeding 0.034 correlated with an ICP below 20 mm Hg (P=.005). Based on this finding, they adopted 0.034 as a threshold to indicate elevated ICP, achieving a predictive accuracy of approximately 67%, indicating that further refinement is needed. Shan et al [113] investigated 3 predictive models using different types of input features: a HU model using HU-related imaging features, an MLS model relying on MLS data, and a clinical expertise (CE) model correlating clinical judgment with the actual ICP value. These models were evaluated for both binary classifications, indicating an increased ICP with a yes or no, and multiclass classification for ICP thresholds ≤22, 23‐29, and ≥30 mm Hg. For the binary classification task, the HU model outperformed the others, achieving an accuracy of 81% and an F₁-score of 0.85. In contrast, the MLS and CE models had lower accuracies of 61.7% and 63.83%, respectively. When extended to the more complex multiclass task, all models experienced performance declines. The HU model’s accuracy dropped to 61.7%, and the MLS model to 40.43%. Most notably, the CE model failed to provide reliable classifications. This decline likely reflects difficulties in distinguishing subtle differences in imaging features that correspond to different increased ICP severities. Li et al [114] introduced a radiomics-based approach that leverages quantitative feature extraction from CT scans. By using PyRadiomics, the study extracted 18 first-order and 40 second-order features. First-order features capture basic statistics of voxel intensities within a RoI, reflecting the overall brightness or density of brain tissue or CSF, including mean, SD, and skewness. In contrast, second-order features describe the spatial relationships between pixel intensities, commonly known as texture features, and are derived using the gray-level co-occurrence matrices or the gray-level run-length matrix. Feature selection was performed using the least absolute shrinkage and selection operator, and 3 models were developed: a clinical features model, a first-order model, and a second-order model. The SO model yielded the best performance, with an accuracy of 80%, an F₁-score of 0.83, and an AUC-ROC of 0.81. This finding highlights the potential of incorporating higher-order radiomic information to enhance predictive accuracy. Table S7 in Multimedia Appendix 3 [112-114] presents the details of the included studies.

Based on our survey, we identified numerous challenges in the current research, limiting AI for TBI application in real-world settings. First, current AI models are designed to detect isolated radiological findings such as ICH, MLS, and elevated ICP. While these targeted approaches have demonstrated high accuracy for individual tasks, they fail to capture the broader heterogeneity of TBI presentations. For instance, increased ICP is typically not an isolated finding, but often results from a combination of factors such as brain lesions, MLS, and compression of the basal cisterns. However, current studies rarely consider these interrelated indicators. This narrow, fragmented approach limits diagnostic performance in real-world settings, where multiple abnormalities commonly occur together. To address these limitations, future research should aim to develop integrated AI systems that combine multiple models into a unified diagnostic pipeline capable of recognizing the full spectrum of TBI-related abnormalities. One promising direction is the use of hierarchical AI frameworks, which incorporate specialized modules for different lesion types and severity levels. Such frameworks could be more context-aware decision support systems.

Second, clinical validation of these AI models is critically lacking. Among the reviewed literature, only one study [106] conducted formal validation involving health care professionals. In the context of this review, “formal validation involving health care professionals” means having clinicians directly involved in validating the performance of the models by actively evaluating the model’s outputs or having independent evaluation of the model’s output against targeted assessments made by clinical experts. Without such targeted clinical validation, model performance metrics, however strong, remain theoretical. Moreover, translation of AI-based TBI assessment tools into clinical practice requires substantially more evidence than is currently available in most studies. Regulatory approval pathways for AI-based diagnostic systems, such as Software as a Medical Device, typically require prospective validation, demonstration of generalizability across sites and scanners, clearly defined clinical use cases, and assessment of human-AI interaction within clinical workflows. The limitation in external clinical validation highlights a gap between methodological development and regulatory or clinical readiness. Addressing these requirements will be essential for future studies aiming to move beyond proof-of-concept toward deployable clinical systems. Thus, it is essential that future AI development be conducted in close collaboration with clinicians to ensure that AI tools not only have consistent technical performance but also have alignment with clinical expectations.

Third, the condition of the basal cisterns such as normal, compressed, or absent is a known prognostic indicator in TBI [115-117] but remains underexplored in AI literature. This gap likely exists due to the subjective nature of interpreting basal cistern conditions, making it difficult to establish clear criteria for training ML models. Similarly, AI-based ICP estimation is another underexplored area, with only 2.9% of the reviewed studies addressing it. Current methods remain insufficiently robust for clinical use. Future studies should investigate the potential for integrating multimodal data, combining imaging and physiological signals to improve the accuracy of ICP estimation.

Furthermore, there is a notable challenge in validating and comparing these AI systems. This is primarily due to the limited availability of publicly accessible benchmark datasets for TBI, especially extensive annotated datasets from clinical experts. Most studies reviewed in this paper relied on private datasets with limited access, which impedes efforts to benchmark segmentation algorithms with consistent training and testing data. The PhysioNet ICH dataset, which includes CT scans from 82 patients with TBI with pixel-level annotations, is currently the only publicly available resource for segmentation tasks. However, the small size of this dataset poses a risk of overfitting when training complex DL models with millions of parameters. To advance the field, there is a pressing need for multi-institutional collaborations to build large, diverse, and publicly available datasets with standardized annotations by expert clinicians. Such resources would not only enable fair comparison of algorithms but also foster the development of clinically viable models.

In addition to the need for larger datasets, AI models for TBI assessments are currently limited by their dependence on supervised learning techniques, which require large amounts of annotated data. This is particularly challenging in the medical domain, where TBI lesion annotations must be performed by clinicians with specialized knowledge. Furthermore, given CT scans represent 3D volumetric data, the annotation process is time-consuming and complex. To alleviate this burden, future research should explore alternative learning strategies, such as semi-supervised or self-supervised learning techniques, which can reduce the reliance on extensive annotated datasets. The study of Yao et al [118] is one of the few attempts to address this issue.

Finally, one of the most significant limitations of AI models in health care applications is their “black box” nature, which limits transparency and interpretability. To improve clinical acceptance and facilitate integration into TBI assessment, AI models should ideally be interpretable and aligned with clinical domain knowledge. Rule-based models, particularly those using fuzzy neural networks (FNN), are a promising solution. FNNs allow the extraction of interpretable “if-then” rules, which can be aligned with CE. This approach not only improves model performance but also reduces the need for large volumes of training data. Additionally, FNNs offer the flexibility to extract meaningful rules even when some variables are missing, making them particularly suitable for real-world clinical environments where data may be incomplete.

This review has several limitations. First, as a systematized review, it lacks prospective protocol registration, duplicate independent screening, and formal risk-of-bias assessment using standardized instruments. Furthermore, our highly aggressive exclusion threshold during the initial title and abstract screening phase, which resulted in zero exclusions at the full-text review stage, carries an inherent risk of selection bias. Relevant studies that failed to explicitly report specific inclusion criteria (such as imaging modality or patient age) in their abstracts may have been prematurely excluded. Although we qualitatively considered key sources of bias such as dataset size, single-center versus multicenter design, annotation practices, and validation strategies, our ability to systematically quantify study quality was limited, which may increase the risk of selection and reporting bias.

Second, the evidence base is constrained by heterogeneous study designs, outcome definitions, and reporting practices, which precluded meta-analysis and may introduce selective reporting and publication bias; for example, negative or inconclusive AI studies may be underrepresented in the published literature. Consequently, the findings should not be interpreted as definitive estimates of diagnostic accuracy or clinical effectiveness. Instead, the primary contribution of this review is to provide a structured mapping of AI tasks, methodological approaches, datasets, and evaluation practices used in CT-based TBI assessment. The review is intended to highlight methodological trends, recurring limitations, and gaps in validation and reporting, rather than to support direct clinical adoption or comparative performance claims between models.

Third, we restricted inclusion to English-language, peer-reviewed papers indexed in 4 major databases and did not systematically search gray literature, potentially missing relevant work. The review was restricted to English-language publications to ensure reliable interpretation of technical and methodological details, as incomplete or inaccurate translation could compromise assessment consistency. Moreover, peer-reviewed publications were prioritized to reduce the risk of incomplete, preliminary, or subsequently revised methodological claims.

Fourth, despite updating the literature search to include recent publications, the rapid pace of research in AI for medical imaging means that newly emerging methods may continue to appear after the search cutoff.

Finally, although multimodal AI approaches integrating CT imaging with clinical or physiological data are increasingly explored, they were excluded from this review by design to maintain a focused analysis of imaging-specific methodologies. Such approaches address distinct methodological and clinical questions related to outcome prediction rather than isolated imaging-based abnormality detection. Therefore, the lack of multimodal analysis in this study represents a limitation of our review scope, rather than a deficit in the broader scientific field. A future dedicated review focusing exclusively on multimodal TBI AI systems, as well as distinct reviews tailored to pediatric neurotrauma, are critically needed to complement these findings.

We have highlighted the progress made in the application of AI for TBI assessment on CT scans. 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 [119], as well as 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 labeling data. Furthermore, ensuring transparency and interpretability of AI models is essential for successful integration into clinical practice and clinical acceptance. Finally, clinical validation with health care 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.