A systematic review highlights that AI model architecture, more than data or imaging specifics, drives improved tumor segmentation in brain MRI scans.
Key Details
- 1Researchers from the University of Auckland and affiliated institutes reviewed 34 studies on AI-based meningioma segmentation using MRI (2020-2025).
- 2Advanced AI architectures, such as DeepLabV3+, achieved the highest Dice scores (up to 0.98), indicating superior delineation performance.
- 3Model performance tiers were identified: Tier 1 (Dice 0.9-1, hybrid/advanced models), Tier 2 (Dice 0.8-0.9, e.g. 2D U-Net), Tier 3 (Dice <0.82, lightweight models).
- 4Data and imaging modality improvements contributed less to performance than advancements in AI architectures.
- 5Challenges remain, including underperformance for small tumors (<3 mL), high computational demands, and limited generalizability across hospitals.
- 6Most studies used contrast-enhanced T1 MRI; real-world deployment still faces workflow and infrastructure hurdles.
Why It Matters
This review provides critical evidence that prioritizing advanced AI architectures is key for accurate brain tumor segmentation in MRI, advancing radiology workflows and research focus. Understanding the limitations will help guide development of more deployable and generalizable models in clinical practice.
Source
EurekAlert
Related News
•EurekAlert
AI-Enhanced Imaging Breaks Barriers in Complex Media for Biomedical Use
Researchers developed a physics-based machine learning imaging system that enhances visibility through complex media, holding promise for biomedical imaging.
•EurekAlert
AI Model Accelerates Accurate Dose Prediction for CyberKnife Radiotherapy
A physics- and spatially informed diffusion model accelerates Monte Carlo–level dose estimation for CyberKnife radiotherapy using planning CT and anatomical data.
•EurekAlert
New G-AUDIT Tool Detects Hidden Bias in Medical AI Training Data
Johns Hopkins and FDA researchers unveil G-AUDIT, a tool to identify hidden biases in medical AI datasets before model training.