Before you write a single line of code, you need to understand why this technology matters. This week, you’ll learn how AI-powered brain tumor segmentation is being used in real hospitals to help real patients — from surgical planning to tracking whether treatment is working.
Every year, hundreds of thousands of people worldwide are diagnosed with brain tumors. Treating them requires incredibly precise information: where exactly is the tumor? How big is it? Which parts are actively growing? Is the treatment shrinking it?
Doctors answer these questions using MRI scans. But here’s the problem: a single brain MRI produces a 3D volume made of hundreds of individual image slices, across multiple imaging sequences (T1, T1ce, T2, FLAIR). A neuroradiologist must manually trace the tumor boundaries on every relevant slice, identifying different sub-regions — a process called segmentation. This can take 30 minutes to several hours per patient, is highly subjective, and varies significantly between experts.
This is where AI comes in. Deep learning models — particularly architectures like U-Net and nnU-Net — can perform this segmentation automatically in seconds, with accuracy that matches or exceeds human experts.
Neuro-oncology is the branch of medicine focused on cancers of the brain and nervous system. As someone building segmentation models, you don’t need to become a doctor, but you do need to understand the basics of what you’re looking at and why it matters clinically.
Gliomas are the most common primary brain tumors, arising from glial cells that support neurons. They’re classified by the World Health Organization (WHO) into grades 1–4, with grade 4 (glioblastoma / GBM) being the most aggressive. Glioblastomas have a median survival of roughly 15 months even with treatment. BraTS has historically focused on glioma segmentation, particularly distinguishing the enhancing tumor core, necrotic regions, and surrounding edema.
These are cancers that started elsewhere in the body (lung, breast, melanoma, etc.) and spread to the brain. They’re actually the most commonly diagnosed brain tumors in adults. Patients often have multiple lesions of varying sizes, making manual segmentation extremely tedious. This is where AI tools are especially impactful — and it’s the focus of the BraTS-METS 2025 challenge.
Tumors arising from the meninges (the membranes surrounding the brain). Usually benign and slow-growing, but they can compress brain tissue and require surgery or radiation. BraTS 2025 includes dedicated meningioma tasks (Tasks 2–3).
Brain tumors are the leading cause of cancer death in children. They differ significantly from adult tumors in location, biology, and imaging appearance. nnU-Net has been validated on pediatric tumors, achieving Dice scores of 0.90 for whole tumor, though performance varies by sub-region. BraTS-PED (Task 6) addresses this population.
Understanding where AI fits requires knowing the clinical workflow:
Patient presents with symptoms (headaches, seizures, neurological deficits). MRI scans are acquired with multiple sequences (T1, T1ce, T2, FLAIR). A neuroradiologist interprets the images and identifies the tumor.
The tumor’s size, location, and sub-regions are assessed. WHO grading may be estimated from imaging (later confirmed by biopsy). Segmentation provides the volumetric measurements needed for treatment decisions.
If surgery is indicated, precise 3D models of the tumor and surrounding anatomy are needed for navigation. AI segmentation can produce these models faster and more consistently than manual methods (mean 1,139 seconds vs 2,851 seconds for traditional approaches).
Radiation oncologists need to delineate the gross tumor volume (GTV) and clinical target volume (CTV). AI-assisted contouring improves inter-reader agreement (Dice from 0.86 to 0.90) and is especially helpful for less-experienced physicians.
After surgery, radiation, and/or chemotherapy, patients undergo regular follow-up MRI scans. AI-based volumetric tracking enables objective, reproducible measurement of whether the tumor is responding, stable, or progressing.
Before brain surgery, neurosurgeons need detailed 3D models showing the tumor relative to critical structures (motor cortex, language areas, blood vessels). AI segmentation can generate these automatically from MRI in under 20 minutes, compared to nearly an hour with traditional mesh-growing algorithms. During surgery itself, these segmentations feed into neuronavigation systems that guide the surgeon in real-time.
For brain metastases treatment with focused radiation, every lesion must be precisely contoured. A randomized multi-reader study showed that AI assistance improved lesion detection from 82.6% to 91.3% sensitivity, while providing a median 30.8% time savings. The benefit was greatest for less-experienced physicians, helping close the expertise gap.
This is one of the most impactful clinical applications. After treatment, doctors need to determine whether the tumor is responding. The RANO criteria (Response Assessment in Neuro-Oncology) define standardized rules for this assessment.
A landmark study in The Lancet Oncology by Kickingereder et al. (2019) demonstrated this powerfully. Analyzing 532 patients from 34 institutions in the EORTC-26101 trial, they showed that AI-based volumetric response assessment was significantly more reliable than manual RANO assessment:
For brain metastases specifically, AI models have achieved 100% accuracy in predicting RANO-BM criteria compared to expert assessment, demonstrating that automated longitudinal tracking is clinically viable.
One of the most exciting developments is embedding AI directly into the tools radiologists already use. A system at Yale New Haven Health integrated a deep learning algorithm directly into their Visage PACS diagnostic workstation, achieving segmentation in 4 seconds with radiomic feature extraction in under 6 seconds. The AI runs before the radiologist even opens the study — they just verify the result.
A natural question: how does AI actually compare to trained neuroradiologists? The answer is nuanced.
A systematic review and meta-analysis reported an overall Dice score of 0.84 (95% CI: 0.82–0.86) for ML-based glioma segmentation. State-of-the-art models on the BraTS 2021 benchmark achieve Dice scores of 0.936 for whole tumor, 0.921 for tumor core, and 0.872 for enhancing tumor — which is within or above the range of human inter-rater agreement (0.74–0.85).
Round, well-demarcated tumors: AI correctly includes necrosis and contrast-enhanced regions in 97–100% of cases (vs 73.68% for manual segmentation). Consistency: AI produces identical results every time — no inter-rater variability. Speed: Seconds vs minutes or hours. Multi-lesion counting: For patients with many brain metastases, AI dramatically reduces the tedium and missed lesions.
Complex, infiltrative tumors: For tumors with irregular borders and diffuse infiltration, manual segmentation outperforms automated methods (66.67% vs 37.50–50% accuracy). Clinical context: Radiologists integrate patient history, prior scans, and clinical knowledge that current AI models don’t have access to. Post-treatment changes: Distinguishing true progression from pseudoprogression (treatment-related changes that mimic tumor growth) remains challenging for AI.
Several AI tools have moved beyond research into clinical practice:
An FDA-cleared deep learning algorithm for automated detection and contouring of metastatic brain tumors in stereotactic radiosurgery. Validated on 100 patients with 435 brain metastases: 89.3% overall sensitivity, reaching 99% for lesions ≥10mm. Mean false positive rate of just 0.72 tumors per case. Performance was consistent across demographic and clinical subgroups.
An open-source platform that makes BraTS challenge algorithms available for clinical use. Three components: BraTS Preprocessor (data standardization), BraTS Segmentor (algorithm orchestration), and BraTS Fusionator (consensus segmentation from multiple models using majority voting). This is how research algorithms get translated into usable tools.
A novel system embedding a UNETR deep learning model directly into the Visage 7 diagnostic workstation at Yale New Haven Health. Achieves median Dice of 86% in 4 seconds, with automatic radiomic feature extraction. Represents the future of seamless clinical integration.
Despite impressive results, widespread clinical adoption faces significant hurdles. Understanding these challenges is essential for anyone entering this field.
This is the biggest technical challenge. Models trained at one hospital often perform worse at another due to domain shift — differences in MRI scanners, imaging protocols, patient populations, and image quality. Cross-institutional studies show Dice score drops from 0.72–0.76 to 0.59–0.68 when models are tested at new institutions. BraTS dataset performance often overestimates real-world performance because BraTS data is heavily curated and standardized.
A 2026 study found that segmentation models trained on demographically homogeneous datasets exhibited the most bias. Models trained exclusively on one demographic group had lower Dice scores and showed biases related to age and other characteristics. The key finding: greater demographic diversity in training data reduced bias even without increasing dataset size.
Getting AI tools approved for clinical use requires navigating FDA (US) and CE marking (Europe) pathways. The AI-RAPNO consensus identified key challenges: variability in imaging protocols, scarce annotated datasets for rare tumor types, lack of standardized validation frameworks, and ethical considerations around transparency and patient consent.
Deep learning models are difficult to interpret — they produce results but can’t explain their reasoning. This erodes clinical trust, especially in high-stakes medical decisions. Newer architectures are incorporating attention mechanisms and clinical knowledge to improve interpretability, but this remains an active research area.
Even when the AI works, deploying it in a hospital setting requires: IT infrastructure for GPU computing, interoperability with existing PACS and EMR systems, data privacy compliance (HIPAA/GDPR), trained staff to maintain the system, and reimbursement models to justify the cost.
The BraTS Challenge has been instrumental in everything described above. Since 2012, it has provided standardized datasets, evaluation metrics, and a competitive environment that pushed segmentation performance from research curiosity to clinical viability.
Key contributions: dataset growth from 65 cases (2012) to over 1,251 cases (2021) and expanding further; standardized preprocessing protocols; expansion from gliomas to metastases, meningiomas, pediatric tumors, and sub-Saharan African populations; and establishment of the Dice/Hausdorff/NSD evaluation framework used across the field.
The Federated Tumor Segmentation (FeTS) Challenge extended this further by allowing 32 institutions to collaboratively train models without sharing patient data — addressing privacy concerns while improving generalizability.
The upcoming BraTS 2025 Lighthouse Challenge continues to expand across additional clinical needs, including post-treatment scenarios and integration with genomic data.