Abstract
Background Brain tumors present significant diagnostic challenges because early and accurate identification is essential for treatment planning and patient outcomes, while conventional diagnostic processes rely heavily on expert interpretation of complex medical images. Magnetic resonance imaging (MRI) remains central to brain tumor diagnosis, but manual evaluation can be time-consuming, cognitively demanding, and subject to inter-observer variability. Recent advances in artificial intelligence (AI), particularly machine learning, deep learning, computer vision, and explainable AI (XAI), have created opportunities to automate tumor detection, segmentation, classification, and characterization. However, translation of high-performing AI models from controlled datasets into clinical practice requires consideration of more than algorithmic accuracy. From a human factors perspective, effective AI must support clinicians’ cognitive work, provide understandable information, reduce rather than redistribute workload, integrate appropriately into clinical workflows, and enable clinicians to maintain meaningful oversight of diagnostic decisions. Objective This study synthesizes recent literature on AI applications for brain tumor diagnosis to characterize both the technological capabilities of current approaches and the human factors issues affecting their potential clinical adoption. Particular attention was given to how AI may influence diagnostic workload, decision-making, clinical confidence, interpretability, and human-AI collaboration, as well as the barriers that remain between high algorithmic performance and effective clinical implementation. Methods A structured synthesis was conducted using 15 publications examining AI applications related to brain tumor diagnosis and management, primarily published between 2022 and 2026. The reviewed literature included empirical model-development studies and broader reviews addressing AI-enabled brain tumor detection, classification, segmentation, imaging analysis, radiomics and radiogenomics, biomarker identification, multimodal diagnosis, and emerging non-invasive diagnostic technologies. For each publication, information was systematically extracted regarding the healthcare problem addressed, clinical application, AI technique, data source, reported findings and benefits, limitations, future research recommendations, and implications for healthcare delivery. The synthesis emphasized recurring themes relevant to human factors and ergonomics, including diagnostic accuracy, speed of decision-making, clinician workload, interpretability and explainability, trust, clinical decision support, human error, workflow integration, data quality, generalizability, computational requirements, privacy, and human acceptance. Because studies used heterogeneous datasets, tasks, architectures, and performance measures, reported accuracy values were interpreted descriptively rather than directly compared or meta-analyzed. Results The literature demonstrates substantial technical progress in AI-supported brain tumor diagnosis. Convolutional neural networks (CNNs), transfer-learning architectures, EfficientNet variants, Vision Transformers, ensemble models, YOLO-based detection, segmentation networks, and hybrid architectures were commonly applied to MRI data. Several studies classified MRI images into glioma, meningioma, pituitary tumor, and no-tumor categories, while others focused on binary tumor detection, automated segmentation, localization, or multimodal diagnosis. Reported performance was generally high, with several models achieving classification accuracy above 95% under experimental conditions. Hybrid approaches combining complementary architectures frequently outperformed individual models by capturing both local tissue characteristics and broader spatial relationships. AI applications also extended beyond conventional image classification. Radiomics and radiogenomics approaches demonstrated the potential to infer tumor characteristics and clinically meaningful molecular information non-invasively. AI-driven biomarker discovery integrated imaging, genomic, transcriptomic, proteomic, and clinical information to support earlier diagnosis, prognosis, treatment-response prediction, and personalized oncology. Another emerging approach combined machine learning with a highly sensitive terahertz metasurface biosensor to explore liquid-biopsy-based tumor detection, illustrating the potential for rapid and less invasive diagnostic pathways. Three-dimensional reconstruction and segmentation-based approaches further demonstrated how AI could support spatial localization, treatment planning, and preoperative assessment. Across these technological approaches, a central human factors benefit was the potential reduction of clinician workload. Automated segmentation, measurement, detection, and classification can reduce repetitive image-review tasks and provide radiologists and neuro-oncology teams with a rapid second opinion. The reviewed literature frequently described AI as a clinical decision-support mechanism rather than a replacement for clinicians. By identifying subtle abnormalities, prioritizing suspicious regions, distinguishing tumor categories, and generating consistent quantitative assessments, AI may reduce diagnostic delays, false-positive and false-negative findings, and variability associated with manual interpretation. However, improved algorithmic performance did not necessarily correspond to clinical readiness. Explainability emerged as one of the most prominent challenges. Many high-performing deep learning systems operate as “black boxes,” creating difficulty for clinicians who must evaluate why a particular diagnosis was generated. Several studies addressed this problem using Grad-CAM, LIME, segmentation maps, or other visualization mechanisms that highlight image regions contributing to a prediction. These approaches can transform AI output from a simple classification into visually inspectable evidence, allowing clinicians to compare model reasoning with their own interpretation. Thus, explainability appears to function not only as a technical feature but also as a human factors mechanism supporting situation awareness, appropriate trust, verification, and shared decision-making between clinicians and AI. Generalizability was another consistent limitation. Many systems were developed using public or single-source MRI datasets, including commonly used benchmark or Kaggle datasets. Differences in scanner characteristics, imaging protocols, patient populations, tumor prevalence, image quality, and institutional practices may substantially influence performance when algorithms are transferred between healthcare settings. Consequently, several studies recommended validation using larger, diverse, multi-center clinical datasets. Data imbalance and limited representation of rare tumor types further raise concerns about whether high average accuracy masks clinically important failure modes for specific patients or diagnoses. Computational and workflow constraints also emerged. Hybrid CNN-transformer architectures, ensemble approaches, 3D reconstruction, and real-time explainability can impose significant memory and processing demands, potentially limiting implementation in resource-constrained settings. In contrast, lightweight models demonstrated that somewhat lower benchmark performance may be accompanied by extremely rapid execution and easier deployment. This creates an important human factors and systems-engineering tradeoff: the most accurate algorithm in an experimental environment may not necessarily be the most useful system within a clinical workflow. Finally, privacy, security, algorithmic bias, regulation, and human acceptance were repeatedly identified as implementation concerns. These challenges become more prominent when systems integrate imaging with clinical records, multi-omic information, cloud infrastructure, or distributed data sources. Successful implementation therefore requires both technical validation and attention to how information is presented, how clinicians interact with AI recommendations, when human review is required, how uncertainty is communicated, and how accountability is distributed between clinicians and automated systems. Discussion and Conclusion Collectively, the literature suggests that AI has substantial potential to improve brain tumor diagnosis by increasing diagnostic accuracy, accelerating image analysis, supporting earlier detection, reducing repetitive workload, enabling non-invasive characterization, and assisting personalized treatment planning. Nevertheless, the major barrier to clinical translation is increasingly not simply whether AI can classify an MRI image accurately. Rather, the challenge is designing a sociotechnical system in which clinicians can understand, evaluate, appropriately trust, and effectively use AI-generated information. Future research should therefore move beyond performance-centered model development toward human-centered clinical evaluation. Multi-institutional validation should be accompanied by usability testing, workflow analysis, assessment of clinician cognitive workload, examination of appropriate reliance and trust, and evaluation of how explainability interfaces influence diagnostic decisions. Studies should also investigate AI performance under realistic conditions, including ambiguous cases, rare tumors, imperfect imaging, competing clinical demands, and disagreement between clinicians and algorithms. Human factors methods can play a critical role in determining where AI should enter the diagnostic workflow, what information should be displayed, how uncertainty should be communicated, and how human oversight should be maintained. The next generation of AI-enabled brain tumor diagnosis should therefore be evaluated not only by “How accurate is the model?” but also by “How effectively does the human-AI system support safe, efficient, interpretable, and clinically meaningful decision-making?” This shift from algorithm-centered evaluation toward human-centered system design represents an important opportunity for human factors and ergonomics to guide the responsible translation of AI into neuro-oncology.
Whitepaper forthcoming.