Deep Learning Applications In Automating Brain Tumor Segmentation On MRI: A Systematic Review Of Clinical Performance

Abstract

Background: Primary and metastatic brain tumors constitute a profound global health challenge, characterized by high morbidity and mortality rates. In 2023 alone, it was estimated that 26,940 new malignant brain tumors would be diagnosed in the United States, with glioblastoma accounting for 50.1% of these malignancies. The condition imposes a severe burden on populations globally, with age-standardized incidence rates showing disparities between genders and geographic regions. The current standard of care for diagnosis, treatment planning, and response assessment relies heavily on Magnetic Resonance Imaging (MRI). Specifically, the precise delineation or segmentation of tumor boundaries is critical for radiotherapy planning and surgical navigation. However, the conventional intervention—manual segmentation by radiologists—is fraught with limitations. It is a labor-intensive, time-consuming process subject to significant inter-observer and intra-observer variability, which can compromise the accuracy of therapeutic delivery. In response to these challenges, Deep Learning (DL), particularly Convolutional Neural Networks (CNNs) and transformer-based architectures, has emerged as a promising alternative intervention. These automated systems offer the potential to standardize quantification and drastically reduce workflow time while maintaining expert-level accuracy.

Objective: The primary aim of this systematic review is to comprehensively and systematically compare the clinical effectiveness of Deep Learning-based automated segmentation (Intervention 1) versus manual segmentation by clinical experts (Intervention 2). The review specifically evaluates geometric accuracy, time efficiency, and clinical utility across diverse patient populations with gliomas, meningiomas, and brain metastases.

Methods: A systematic review was conducted in strict adherence to the PRISMA 2020 guidelines. A comprehensive search was performed across major medical and technical databases, including PubMed, Scopus, IEEE Xplore, and Web of Science, covering the period from 2015 to 2024. The study selection was guided by the PICO framework: Population (patients with brain tumors on MRI), Intervention (Deep Learning models), Comparison (Manual segmentation), and Outcomes (Dice Similarity Coefficient, Hausdorff Distance, processing time). The risk of bias in included prediction model studies was rigorously assessed using the PROBAST tool.

Results: The search identified a substantial corpus of evidence, from which key studies meeting strict inclusion criteria were analyzed. The synthesis of data reveals that Deep Learning models, particularly the nnU-Net and hybrid transformer architectures, demonstrate non-inferiority to manual segmentation. For adult gliomas, hybrid models achieved Dice Similarity Coefficients (DSC) exceeding 0.90 for whole tumor segmentation. In pediatric cohorts, nnU-Net outperformed older architectures like DeepMedic, achieving a mean DSC of 0.90 versus 0.82. For meningiomas, DL models demonstrated a DSC of 0.91, statistically equivalent to the inter-reader variability of human experts. Most significantly, DL integration reduced segmentation time by approximately 98%, cutting the process from an average of 20 minutes to under 10 seconds per case.

Conclusion: Deep learning algorithms have reached a level of maturity where they offer geometric accuracy comparable to human experts while providing superior time efficiency. The evidence suggests that DL can effectively alleviate the radiological burden, enabling rapid adaptive radiotherapy and standardized longitudinal monitoring. However, significant barriers regarding generalizability to external datasets and integration into clinical workflows persist. Future research must prioritize multi-institutional validation and explainable AI to ensure safe clinical adoption.