Validated Swin Transformer-Based Deep Learning Pipeline with Cross-Validation and McNemar’s Test for Multi-Class Hemorrhage Classification in Traumatic Brain Injury CT Scans

I.INTRODUCTION

Traumatic brain injury (TBI) is a critical population health problem that is estimated to impact millions of individuals per year and is one of the most common causes of death and disability in the global population. Globally, a total of more than 69 million people experience TBI annually, and the distribution of the issue is incredibly skewed. TBIs in developed countries can be linked to sport injury, falls among the elderly, and road traffic accidents, whereas road traffic accidents continue to be the leading cause of TBI prevalence in low- and middle-income countries. It is important to note that TBI has enormous financial costs in the short term (direct clinical consequences) and the long term (indirect consequences) affecting the family, healthcare delivery systems, and the economy through cognitive, physical, and psychosocial effects. Chronic impairments in attention, memory, and motor coordination are common in survivors of TBI, which decreases their quality of life and leads to chronic dependence. These dimensions make the accurate diagnosis, timely treatment, and successful classification of TBI subtypes essential not only to the fate of individual sufferers but also to the reduction of the general impact of the disorder on society.

Intracranial Hemorrhage (ICH) is one of the most clinically significant consequences of TBI and occurs when blood gathers in or around the brain due to traumatic injuries to blood vessels. ICH does not exist as an entity, but it is a continuum of various categories of hemorrhages, some of which include epidural, subdural, subarachnoid, intraparenchymal, and intraventricular hemorrhages. Each of these subtypes implies a different clinical implication. When it comes to skull fractures, the epidural hemorrhages can grow and require urgent neurosurgical intervention. Subdural hemorrhages, in turn, may proceed at a slower pace, yet they are also associated with a high mortality rate in the case of failure to treat them. Subarachnoid bleeding is associated with the risk of vasospasm and delayed ischemia, and intraventricular bleeding could lead to the blockage of cerebrospinal fluid channels and consequently to hydrocephalus. The intraparenchymal hemorrhages that occur in brain tissues may cause devastating focal deficits depending on their localization. In light of such multiplicity, the appropriate categorization of ICH subtypes is critical to selecting the appropriate clinical pathway and prioritizing interventions in acute care.

Non-contrast computed tomography (CT) is the standard imaging modality in modern clinical practice, since it is employed to detect the existence of intracranial hemorrhage within the shortest time possible. CT is associated with a number of advantages: it is standard, it is less expensive when compared to Magnetic Resonance Imaging (MRI), and the images they depict can be delivered very fast when faced with an emergency in an acute setting. Acute blood is very sensitive to CT and, thus, it is very useful in trauma treatment where time is a very important consideration. Reliance on CT is problematic in itself, however. First, radiologist interpretation is never a quick process and can be prone to inconsistency as it can be influenced by human factors. In busy emergency departments where the radiologist must attend to huge caseloads, there is an increased risk of delayed diagnosis. Second, inter-observer variability exists, especially with the mild or marginal cases. In other words, a small subarachnoid hemorrhage may be missed during one examination and identified during another. Finally, the availability of qualified and prompt evaluation is restricted due to a large gap in the number of trained neuroradiologists, particularly in low-resource settings. These weaknesses indicate that there is an acute necessity to create augmented diagnostic devices that can help radiologists improve the speed and accuracy of hemorrhage classification.

In recent years, machine learning (ML) and artificial intelligence (AI) have become the disruptive technologies in the medical imaging industry, which can answer the majority of these issues. AI models have already been shown to be useful in other domains of TBI management, such as image processing via computer and outcome prediction. They can be explainable ML models such as E-TBI that can provide interpretable patient post-TBI outcome predictions that guarantee the decision-making process is both transparent and clinically meaningful [1]. A different study also proved that ML can be effective in forecasting when emergency neurosurgery is required within 24 hours of severe or moderate TBI and offer clinicians real-time decision support tools in life-threatening scenarios [2]. Biomarker-based diagnostics have also been used in conjunction with computational models where proteins were quantified to determine the extent and course of injury, such as glial fibrillary acidic protein (GFAP) and ubiquitin carboxy-terminal hydrolase L1 (UCH-L1) [3]. These examples represent the scope of AI implementation in TBI and how it could be applied to integrate imaging, clinical, and molecular data and deliver holistic patient care.

The use of deep learning models has brought a revolution to the automated interpretation of CT scans in the imaging field. They have been shown in validation studies to achieve radiologist level in bleeding and injury severity grading [4]. Besides detection, deep learning models have also been applied to other secondary injury mechanisms, such as the role of mitochondrial dysfunction in propagating inflammation and cell damage following TBI [5]. These understandings would extend beyond the short-term diagnosis to help clinicians understand the biological foundations of long-term sequelae. It is worth noting that AI-based solutions have been demonstrated to be applicable in other healthcare settings. It is established that predictive frameworks can be highly generalizable in low-resource and high-income environments, implying that they may help to decrease inequality in access to health worldwide [6]. The additional new trends include non-invasive diagnostic assays, for example, acoustofluidic separation as a means of identifying TBI biomarkers in circulating exosomes [7], and the use of ML models to conduct multi-class imaging classification [8].

Going a step further, researchers have examined the application of AI to pediatric groups, wherein TBI poses distinct diagnostic and management issues. Indicatively, neural network-related solutions have been suggested to identify mild TBI in children, emphasizing the adaptability of AI models to the cohort specializations [9]. Moreover, ML-based mortality prediction models have offered patient prognosis assessment tools and have been applied to intensive care and trauma patients [10]. Collectively, these papers indicate that AI is not limited to one area of TBI research but cuts across imaging, biomarkers, patient outcomes, and prognostic modeling.

In spite of these developments, however, there is still a massive gap in research: no strong frameworks of fine-grained, multi-classification of subtypes of intracranial hemorrhage are available. Most of the published literature has concerned binary classification, for example, hemorrhage versus non-hemorrhage, or outcome prediction, for example, mortality or surgical requirement. Although such methods are clinically helpful, they fail to assist in the more subtle problem of discriminating between subtypes of hemorrhage, each of which has its own prognostic and therapeutic consequences. As an example, it is necessary to distinguish between subarachnoid and intraventricular hemorrhages, as they require different management approaches, but this distinction is poorly studied with automated systems. This is a limitation that explains why there is an urgent need to develop models that transcend binary classification and address the multi-class labeling challenges in CT scans.

To fill this knowledge gap, the current study proposes the Swin Transformer, which is a new Vision Transformer (ViT) architecture that can be optimally trained to achieve a trade-off between computational efficiency and high performance in image classification. Conventional convolutional neural networks (CNNs) have been the dominant medical imaging technology since they can capture local patterns. CNNs are, however, naturally constrained by their fixed receptive fields that restrict their capacity to capture long-range interdependencies and global context. By comparison, ViTs perceive images as a sequence of patches and use self-attention to learn relationships on a whole image. This makes it possible to extract rich features and perform better in context-dependent activities. The Swin Transformer also builds upon the unstructured ViTs by implementing hierarchical windowed self-attention; that is, the calculation of attention is carried out in local windows to minimize computational cost. To maintain that global context, the architecture has used shifted windows which permit information to spread in various parts of the image. This is the best of both worlds: efficient computation and global receptive fields.

The novelty of this study is in its use of the Swin Transformer for the problem of multi-classification of subtypes of intracranial hemorrhage, thereby further extending the binary detection paradigms, which currently dominate the literature. In the following way, we can summarize our contributions. We create a balanced dataset of 60 000 CT images using a 1-to-1 ratio of six classes: epidural, intraparenchymal, intraventricular, subarachnoid, subdural, and normal. This ensures that the model is trained in an equal representation of classes, and this will minimize the bias that natural imbalances in real-world data introduce to the model. Second, we conduct Swin Transformer training on this dataset using a highly developed training program that includes optimization schemes such as AdamW, cosine annealing scheduling, and label smoothing to reach the highest generalization. Third, we conduct stringent statistical validation, including K-fold cross-validation to test stability and the McNemar test to test statistical significance compared to a strong base model (EfficientNetV2-S). Finally, explainability techniques are applied as well, though through Grad-CAM++, to ensure that one of our models bases their decision-making on clinically relevant image regions.

In thus doing, this work establishes a new standard in the multi-class ICH classification, not only in quality but also in the rigor of validation. They demonstrate both state-of-the-art performance and statistically robust, interpretable performance, which are essential in clinical translation. The study is itself a contribution to a gap in the existing literature, since it does not stop at binary detection tasks and fine-grained classification, which are the foundations of future AI-based neurotrauma care devices.

The rest of this paper is organized in the following way. Section II conducts a literature review of pertinent literature in the area of biomarkers, imaging, and ML models and explainable AI. In Section III, the dataset, preprocessing pipeline, Swin Transformer architecture, and the experimental setup are provided. The results are reported in Section IV (quantitative measures, statistical validation, and explanation tools). A discussion of the main contributions, clinical implications, and future research directions is provided in Section V.

II.LITERATURE REVIEW

Using AI in TBI research has evolved on many fronts, since it is a complicated disorder and its clinical presentations can have different manifestations. The first neuroimaging research is on functional and metabolic imaging, and researchers started investigating how advanced imaging technologies can be used to detect the functional neural perturbations. Vedaei et al. [11] reported that resting-state functional MRI (rs-fMRI) and ML methodology could improve the discrimination of chronic mild TBI and therefore indicated the potential of multimodal imaging features to unveil the patterns of injury that are likely to pass through the usual CT scans. These papers have shown the relevance of deep learning to detect the biomarkers of connectivity disruption, which is subtle and challenging to identify through human experts. The other significant limitation in these works, however, is that the more complicated a model is, the less the decision-making process could be understood, which brought up the question of interpretability and clinical acceptability.

This is what catalyzed explainable AI in TBI research. To address this, Ngo et al. [12] suggested the E-TBI system, which combined predictive modeling with interpretability features that allow a clinician to understand and trust the algorithmic decisions made in the system. Consistent with that, the explainable ML employed by Liu et al. [13] in predicting sepsis in TBI patients proves transparency to be an important factor in high-stakes clinical predictive processes. Explainability has become a broader trend in medical AI: performance metrics such as accuracy or recall are not sufficient without interpretability mechanisms that are clinically consistent. Such congruency is especially topical in the context of TBI, where the treatment decisions might be life or death.

In addition to explainability, biological and cellular markers have emerged as the new target interest of the TBI research, which offers a possibility to extend the imaging-only diagnostics. Surrogately, extracellular mitochondria have been shown to mediate post-TBI-induced acute lung injury to interrelate neural trauma and systemic inflammatory pathways [5]. It has been proposed as well that NINJ1-mediated endothelial rupture is a predisposing factor to the escalation of blood–brain barrier destruction after TBI, which also plays a role in accelerating the development of the neurological failure [12]. These findings restate that TBI is systemic and that secondary injuries in peripheral organs and microvascular damage within the brain exacerbate this primary trauma. Subsequent pediatric-based reviews followed this study by Lampros and co-authors [13] who concluded that the biomarkers (GFAP and UCH-L1) had special potential in the early diagnosis of susceptible groups. Additional studies by Kim et al. [14] estimated that AI-enhanced neurocritical care systems would use these biomarkers as part of continuous monitoring pipelines, which feed clinicians with real-time feedback. In the same vein, Yao et al. [15] have performed a systematic review of the prognostic applications of ML in mild TBI and have shown how biomarker-based models will guide individual rehabilitation decisions.

Simultaneously, advances in biosensing technologies have driven progress toward point-of-care diagnostics. Özcan et al. [18] developed a quartz tuning fork (QTF) biosensor for GFAP detection, providing an innovative, mass-sensitive approach for rapid screening. Tan et al. [32] introduced surface-enhanced Raman scattering assays for detecting UCH-L1, while Gao et al. [34] designed an ultrasensitive lateral flow assay for S-100β, a well-established marker of TBI. Legramante et al. [33] further demonstrated that GFAP and UCH-L1 assays could triage mild TBI cases by identifying patients unlikely to require CT scans, thereby reducing unnecessary radiation exposure. Collectively, these studies highlight a translational shift toward bedside diagnostics, though their direct application to hemorrhage subtype classification remains indirect, focusing instead on severity stratification and triage.

Although the biomarker approach has emerged, CT imaging remains the clinical gold standard in the evaluation of acute TBI, especially in the case of intracranial hemorrhage. Several experiments have shown that deep learning on the raw CT data is both possible and more precise. Jiang et al. [4] confirmed the use of a deep learning model to detect TBI and Neuroimaging Radiological Injury Score (NIRIS) grading in a multi-reader environment with promising results and challenges in identifying small lesions. By combining CT imaging with bioclinical markers at presentation, Mekkodathil et. al. [16] were able to predict mortality better, and it is proposed that the combination of multimodal data streams will help improve outcome estimates. Uparela-Reyes et al. [17] demonstrated the expanded use of AI in TBI imaging through a bibliometric analysis that demonstrated an exponential increase in the number of publications over the last 10 years, as the world began to become interested in the use of computational models. They also pointed out, however, that to date, the majority of AI models have not moved beyond research to routine clinical practice, and that validation, generalizability, and acceptability to clinicians are all problematic.

Another promising area has been segmentation-based methods, which allow to accurately measure the volume of hemorrhage. A CT-based hemorrhage segmentation system is reported by MacIntosh et al. [21] and has been demonstrated to be useful in both stroke and TBI. However, their results also showed limitations in extrapolability to other scanners and imaging protocols, which is a common problem with AI when it comes to radiology. Synthetic reviews by Lin and Yuh [22] and Gudigar et al. [23] summarized the set of computational strategies used to process TBI imaging, such as CNNs, feature-based, and hybrid. The two reviews stressed that AI models are always associated with a high level of performance on internal datasets, and external validation is poorly represented. Wu et al. [24] further applied this trend to outcome prediction in severe TBI and demonstrated that ML performed better as compared to conventional statistical methods to estimate mortality. These initiatives demonstrate the opportunities of AI to supplement radiological knowledge but also warn against overinterpretation because it has not been multi-institutionally validated.

Prognostic modeling has also become a concomitant research area, aiming to forecast secondary outcomes (hematoma progression, systemic complications, and survival). Mohammadzadeh et al. [25] performed a systematic review and meta-analysis demonstrating that ML could be used in relation to predicting hematoma progression, which is one of the most crucial factors to determine its prognosis. Phaphuangwittayakul et al. [26] took this step further and elaborated a deep learning method of multi-lesion quantification, associating the identification of lesions with the quantification of their volumes. The clinical usefulness of computational methods is proven by a meta-analysis by Courville et al. [27], who found ML algorithms to be superior to traditional models at outcome prediction. The scope of this is extended by Peng et al. [28], who forecasted severe acute kidney injury in patients with TBI, and Ge et al. [31], who simulated sepsis-related acute brain injury, both of which explored how systemic complications might be predicted through AI models. Critical comments on biomarker-based prognostication and the problems with reproducibility in neuroimaging studies were made in complementary reviews by Ghaith et al. [29] and Pierre et al. [30], respectively.

Biosensing is an ongoing parallel innovation that continues to expand the possibilities of prognostics. Tan et al. [32] have shown that Raman scattering-based assays can also detect UCH-L1 with high sensitivity, providing the potential to conduct early prognostic triage. In an article by Legramante et al. [33], the authors contrasted the results of their tests, which included GFAP and UCH-L1, with CT exclusion protocols, and emphasized their utility in eliminating unnecessary imaging in patients with mild TBI. Gao et al. [34] reported an enhanced lateral flow assay of S-100 2 that expands the limits of portable, ultrasensitive biomarker detection. Nevertheless, despite these technological advances, Hibi et al. [35] raised some concerns about the preparedness of existing automated CT-based models to be integrated into clinical practice by citing limitations in dataset diversity, algorithm stability, and understandability.

In these diverse directions, a somewhat homogeneous strand has been located: the increasing prominence of hybrid and explainable architectures. Singh and Rakhra [8] have also emphasized the integration of imaging and outcomes into single AI pipes and how this could mediate between diagnostic and prognostic processes. Lampros et al. [13] emphasized that a child population has specific needs since TBI is linked to various diagnostic and treatment problems in this patient group. Kim et al. [14] approximated that in the future, AI-enhanced neurocritical care would be introduced into the world, which will encompass continuous imaging, biomarkers, and clinic data to develop full-scale decision-support systems. Of paramount interest has been the explainability of such innovations. In the works of Ngo et al. [12], Liu et al. [13], and others, the key to clinician trust is transparent models. Gradient-weighted Class Activation Mapping (Grad-CAM) and SHapley Additive exPlanations (SHAP) are only two examples of methods that have been applied to characterize the decisions made by neural networks, so as to align the algorithmic truth with radiological judgment, and to address the black-box prediction/clinical explanation gap [27].

Combined, the literature reviewed indicates significant progress in AI in TBI in the areas of imaging, biomarkers, prognostication, and biosensing. Nevertheless, it also indicates an underlying and long-standing research gap: the multi-classification of subtypes of hemorrhage in CT scans has not been well researched. Current available work is strongly biased in detecting binaries or prognosticating outcomes, and the immediate diagnostic task of distinguishing epidural, subdural, subarachnoid, intraparenchymal, intraventricular, and normal scans is significantly underexploited. Although biomarkers and prognostic models guide secondary management, the basis on which treatment strategies are developed is the acute classification of the subtypes of hemorrhage. In this work, the Swin Transformer is implemented to fill in this gap by deploying a hierarchical ViT architecture to realize effective multi-class hemorrhage classification with stringent statistical validation and explainability. By placing our work in this more general flow of AI research, we are helping to move our field beyond binary and prognostic paradigms, to clinically actionable, fine-grained diagnostic support. [11–35].

Although hybrid CNN–Transformer architectures have shown potential in specialized clinical studies, most publicly available multi-class hemorrhage datasets report comparatively lower performance, typically in the 70–80% range. Achieving an accuracy of 82.02% in this work therefore represents an improvement over several prior transformer- and CNN-based approaches trained under similar data constraints. Higher accuracies reported in hybrid frameworks generally rely on private, richly annotated datasets and complex multi-branch models that limit reproducibility. As the availability of expert annotations improves, future extensions of this work will explore hybrid fusion models to further enhance classification granularity and performance.

Despite the significant progress achieved by CNN-based and hybrid deep learning models in intracranial hemorrhage detection, existing works remain limited in their ability to capture long-range spatial dependencies and global contextual relationships within CT images—factors that are critical for reliable subtype-level classification. Most prior studies also rely on imbalanced datasets, lack statistical validation, or provide limited interpretability, restricting their clinical applicability. Furthermore, only a few recent approaches have explored transformer-based architectures for medical imaging, and these are primarily focused on binary detection tasks rather than comprehensive multi-class categorization. This gap highlights the need for a unified framework that integrates global attention mechanisms, robust validation strategies, and explainable visualization to enhance diagnostic confidence. Building upon these observations, the next section presents the proposed Swin Transformer-based pipeline, detailing the dataset construction, architectural components, training strategies, and evaluation protocols employed in this study.

III.METHODOLOGY AND EXPERIMENT

IV.RESULT AND DISCUSSION

V.CONCLUSION AND FUTURE WORK

In this study, we presented a validated Swin Transformer-based deep learning framework for the multi-class classification of intracranial hemorrhage in TBI CT scans. The hierarchical ViT architecture demonstrated strong capability for modeling both local and global contextual features through windowed and shifted self-attention mechanisms. Using a balanced dataset of 60,000 CT slices generated through structured augmentation of self-collected scans from Shriman Hospital, Jalandhar, the proposed model achieved an accuracy of 82.02%, which surpassed the EfficientNetV2-S baseline. The statistical significance of this improvement was confirmed through McNemar’s test, while 5-fold cross-validation showed consistent stability across folds. Interpretability analysis using Grad-CAM++ further revealed that the model consistently focused on clinically relevant hemorrhagic regions, strengthening its potential clinical value.

Although the accuracy gain over EfficientNetV2-S was modest, the Swin Transformer provided clearer advantages in terms of robustness to imaging perturbations, more localized and clinically meaningful attention maps, and faster inference speed. These characteristics indicated that the model is well suited for real-world emergency triage and decision support.

This work, however, was constrained by its single-center dataset, which may limit variability in scanner types, imaging protocols, and demographic diversity. The dataset was not combined from multiple institutions, and although augmentation enhanced class balance, broader generalization may still require more heterogeneous data sources. To address this, future research will incorporate multi-center clinical datasets, richer radiologist-verified annotations, and hybrid CNN–Transformer fusion strategies to further improve classification granularity and robustness. Overall, the findings demonstrated the potential of transformer-based architectures to enhance automated hemorrhage assessment and support radiologists in time-critical neuroimaging workflows.

Future work will explore scaling to larger multi-center datasets, integrating clinical and biomarker data alongside imaging, and extending the framework with multimodal fusion approaches. Moreover, further research into explainability and uncertainty quantification will be pursued to enhance clinical trust and deployment readiness.