Advanced AI architectures improve meningioma detection on MRI scans

Meningiomas are the most common brain cancers arising from the protective soft tissue cover of the brain called the meninges. Accurately identifying and outlining them on MRI scans is very important for early diagnosis, treatment planning, and prompt follow-up. Doing this manually takes a lot of time, can vary between doctors, and is especially difficult for small or irregular tumors. The advent of artificial intelligence (AI), especially deep learning (DL), has made it possible to perform this automatically, improving speed, consistency, and usability across different hospitals and imaging systems.

Scientists at the University of Auckland, Auckland City Hospital, and Matai Medical Research Institute in New Zealand have reviewed 34 articles dealing with the use of AI for the delineation of meningiomas from MRI scan images published from 2020 to 2025. They have analyzed the current developments in this field, factors that contribute to the development of a good AI model, and how the shortcomings of the current models can be improved to generate a model that can be used with available resources across hospitals around the globe. The results have been published in the journal Brain Network Disorders, made available online on February 13, 2026, and published in Volume 2, Issue 1 on March 26, 2026.

"More advanced AI models have clearly contributed to improved tumor delineations over time, with their performances becoming better and more consistent," says Dr. Hamid Abbasi, the lead author, Senior Research Fellow, and a Principal Investigator in the Auckland Bioengineering Institute (ABI) and Center for Brain Research at the University of Auckland. "But how they improved was surprising. Smarter models matter more than additional data or better scans," he explains. Studies have used MRI images from various repositories like Figshare and BraTS. However, many still relied on custom datasets from hospitals. 'Contrast-enhanced T1 MRI' was found to be the most effective and commonly employed imaging method for detecting meningiomas. However, irrespective of the number, quality, or type of MRI images used to teach the model, better model architecture remained the main driver of better performance.

The accuracy and robustness of segmentation models are commonly evaluated using a metric called 'Dice score', which quantifies the spatial overlap between the predicted segmentation and the ground-truth annotation, and was used here to compare the performance of the AI models included in this review. The models were divided into 3 tiers based on the dice score. The most efficient models with the highest scores between 0.9 and 1 were grouped into 'tier 1'. They were mostly hybrid models with added features and enhanced architectures. Among them, the top model was 'DeepLabV3+' with a score of 0.98, developed based on a large dataset of 3064 images from a database called 'Figshare'. They gave highly accurate results but were computationally very heavy. 'Tier 2' models like '2D U-Net', with a score around 0.8 to 0.9 Dice, offered the best balance of accuracy and efficiency. 'Tier 3' models (e.g., SDAN) were lightweight, designed for speed and low computation with accuracy less than 0.82 Dice.

"Newer models have better tumor delineation capacity, provide more consistent results and faster performance with some reports being generated in as low as 15 seconds, which is not humanly possible," says Nima Sadeghzadeh, the first author.

But few challenges remain. Tumors less than 3 mL are sometimes missed. Also, the high computational demands may not be viable for the resource-constrained settings, yet. Furthermore, the models are still not generalizable across all hospitals. Future research focusing on AI models that work well across different hospitals and datasets, and making them more efficient for real-world clinical settings, will help push for their faster adoption worldwide.