Advancing Brain Tumor Diagnosis Using Deep Learning: A Systematic Review on Glioma Segmentation and Classification on Multiparametric MRI

Technical Summary: Advancing Brain Tumor Diagnosis Using Deep Learning

Problem Statement
Brain tumors, particularly gliomas, represent a lethal and morphologically complex form of cancer requiring precise diagnosis for effective treatment planning and patient monitoring. Gliomas exhibit heterogeneous growth patterns, ranging from low-grade (LGG) to aggressive high-grade (HGG) forms, often with poorly defined boundaries and significant inter-patient variability. Accurate diagnosis relies on two critical steps: segmentation (delineating tumor subregions such as the whole tumor, core tumor, and enhancing tumor) and classification (identifying tumor type or grade). Traditional manual segmentation is time-consuming, prone to inter-observer variability, and requires high expertise. While Magnetic Resonance Imaging (MRI) is the gold standard for non-invasive assessment, the complexity of multiparametric MRI data necessitates advanced computational tools to automate these tasks and improve diagnostic efficiency.

Methodology
This study is a systematic review conducted in accordance with PRISMA guidelines, evaluating Deep Learning (DL) approaches for glioma segmentation and classification published between January 2022 and March 2025.

• Search Strategy: A comprehensive search of PubMed and Scopus databases identified 310 papers. After applying strict inclusion criteria (full-text English, peer-reviewed, DL-based segmentation of gliomas, reporting of performance metrics like Dice Similarity Coefficient [DSC] or Intersection over Union [IoU]), 31 studies were selected for segmentation analysis.
• Classification Subset: Among the 31 studies, 8 were further analyzed for their ability to perform tumor classification (type or grading) alongside segmentation.
• Data Extraction: The review extracted data on study objectives, datasets (e.g., BraTS, BTF, TCGA-LGG), imaging modalities (primarily multiparametric MRI), preprocessing techniques, DL architectures, hyperparameters, and performance metrics.
• Quality Assessment: The Quality Assessment of Diagnostic Accuracy Studies (QUADAS-2) tool was adapted to evaluate the risk of bias in patient selection, index test development, reference standards, and flow/timing.

Key Contributions and Results

• Segmentation Performance:

  • Datasets: The majority of studies utilized publicly available datasets, with BraTS being the most prevalent. The median patient cohort size was 369.
  • Architectures: Convolutional Neural Networks (CNNs), specifically U-Net variants, dominated the field. Advanced models integrating residual blocks, attention mechanisms, and multi-scale feature aggregation consistently outperformed classical U-Nets.
  • Metrics: Classical U-Nets achieved DSC scores around 80%, while advanced models exceeded 90%. The median overall DSC across all studies was 80.73%.
  • Regional Performance: Performance varied by tumor subregion. Whole Tumor (WT) segmentation achieved the highest median DSC (89.20%), followed by Core Tumor (CT) at 83.56%, and Enhancing Tumor (ET) at 80.04%. Models utilizing attention gates and residual connections (e.g., 3D RR-UNet2+, ResMHA-Net) demonstrated superior capability in delineating complex boundaries.
  • Validation: Most studies used hold-out validation; only a few employed cross-validation or external validation, raising questions about generalizability.

• Classification Performance:

  • Tasks: Classification tasks focused on distinguishing tumor types (glioma vs. meningioma vs. pituitary) or grading (HGG vs. LGG).
  • Performance: Classification models reported high accuracy, ranging from 91.3% to 99.4%, with sensitivity and specificity typically above 95%.
  • Approaches: Most studies used CNNs to extract features for classification, often within a multi-task learning framework. One study utilized an XGBoost model with radiomic and segmentation-derived features.
  • Explainability: A significant gap was identified in the application of Explainable AI (XAI). Only one study in the segmentation pool and one in the classification pool applied XAI methods (e.g., Grad-CAM) to interpret model decisions, highlighting a lack of transparency in current pipelines.

• Risk of Bias:

  • The review identified limitations in patient selection, as most studies relied on curated public datasets rather than consecutive or random recruitment, leading to potential selection bias.
  • Many studies lacked detailed reporting on annotation procedures (number of raters, expertise levels) and validation strategies, resulting in unclear or high risks of bias in several domains.

Significance and Claims
The paper asserts that current AI-based pipelines demonstrate strong potential for accelerating the decision-making process in brain tumor diagnosis. The results indicate that DL models can achieve high accuracy and efficiency in both segmenting complex glioma subregions and classifying tumor types or grades.

However, the authors maintain a modest stance regarding clinical deployment. They emphasize that while performance metrics are promising, the lack of standardized imaging protocols, the reliance on public datasets, and the scarcity of external validation limit the immediate generalizability of these models to real-world clinical settings. Furthermore, the paper highlights a critical need for greater model transparency; the infrequent use of explainable AI approaches is identified as a barrier to fostering clinical trust. The authors conclude that future research must prioritize standardized protocols, rigorous external validation, and the integration of explainability frameworks to ensure these tools are robust, reproducible, and trustworthy for clinical practice.