Brain Tumour MRI Segmentation: Annotation for Neuro-Oncology AI

Why Brain Tumour MRI Segmentation Is One of Medicine's Hardest Annotation Problems

Brain tumours affect approximately 308,000 people worldwide annually, with glioblastoma multiforme (GBM) representing the most aggressive primary brain tumour and accounting for roughly 15% of all brain tumours (Central Brain Tumor Registry of the United States, 2023). Median survival for GBM remains approximately 14–16 months despite surgery, radiotherapy, and temozolomide chemotherapy — making AI-assisted early detection and treatment response monitoring a high-priority clinical AI application.

MRI is the imaging standard for brain tumour diagnosis, staging, and follow-up. Unlike CT or chest X-ray, brain MRI for tumour assessment requires four co-registered sequences acquired in a single session: T1-weighted (for anatomy and haemorrhage), T1 contrast-enhanced (T1ce, for active tumour enhancement), T2-weighted (for oedema extent), and FLAIR (for peritumoral infiltration). Each sequence reveals different tumour biology. A neuroradiologist interpreting these scans integrates information across all four simultaneously — a skill that takes years of subspecialty training to develop.

The BraTS (Brain Tumour Segmentation) challenge, running since 2012 and now the standard benchmark for brain tumour AI, has consistently found that top-performing models require training on neuroradiologist-annotated multi-sequence datasets. The BraTS 2021 dataset — 1,251 pre-operative MRI cases annotated by expert neuroradiologists — remains the most widely used public benchmark and the baseline against which new models are validated.

The Four-Sequence Protocol: What Annotators Must Do Simultaneously

Brain tumour MRI segmentation requires annotating across all four co-registered sequences at the same time. Annotating only on T1ce (the most visually obvious sequence) and copying masks to other sequences is a common shortcut that produces systematically biased training data — the oedema boundary visible on FLAIR does not coincide with the enhancing tumour boundary on T1ce.

The three tumour sub-regions annotators must delineate (using the BraTS hierarchy):

• Enhancing tumour (ET): Regions that take up gadolinium contrast on T1ce, indicating active tumour with disrupted blood-brain barrier. This is the highest-priority region for surgical and radiotherapy planning.
• Tumour core (TC): The union of enhancing tumour and necrotic/cystic core. The necrotic core appears as non-enhancing, heterogeneous low T1ce signal within the tumour boundary.
• Whole tumour (WT): The complete tumour including oedema — the full extent of T2/FLAIR hyperintensity surrounding the core. This is typically the largest region and the most variable between annotators at its margins.

Annotators must draw consistent 3D boundaries across every axial slice where each sub-region is visible, then verify consistency in coronal and sagittal planes using multiplanar reconstruction. A glioblastoma tumour that spans 40–80 axial slices at 1mm slice thickness requires coherent 3D masks that do not jump or contract arbitrarily between adjacent slices.

AI Taggers provides specialist MRI annotation services with board-certified neuroradiologist-supervised workflows designed for multi-sequence brain tumour segmentation at the standard required for regulatory submissions.

The Annotator Credential Problem: Why Inter-Reader Agreement Demands Subspecialists

Brain tumour segmentation is one of the areas in medical AI where annotator credential matters most quantifiably. A 2023 study in Neuro-Oncology Advances comparing neuroradiologist and general radiologist annotations on 200 glioblastoma MRI scans found:

• Enhancing tumour Dice: Neuroradiologists 0.87 ± 0.06 vs general radiologists 0.61 ± 0.14
• Whole tumour Dice: Neuroradiologists 0.91 ± 0.04 vs general radiologists 0.74 ± 0.11
• Necrotic core Dice: Neuroradiologists 0.79 ± 0.09 vs general radiologists 0.52 ± 0.18

The performance gap is largest on the necrotic core and on tumour margins — precisely the regions that matter most for surgical planning and radiotherapy target volume delineation. Necrotic core misclassification in training data directly trains models to misidentify pseudo-progression (treatment-induced necrosis) as true tumour recurrence, a clinically consequential error.

For FDA 510(k) or De Novo submissions, the FDA's guidance on AI/ML-based Software as a Medical Device (SaMD) expects annotator credentials to be documented and appropriate to the clinical task. Using general radiologists for brain tumour segmentation is unlikely to survive regulatory scrutiny.

Multi-Reader Adjudication: The Quality Standard for Production Datasets

The BraTS benchmark explicitly requires multi-reader annotation: each case in the BraTS 2021 dataset was annotated independently by three to four neuroradiologists, with consensus labels generated by majority vote or fusion on each voxel. This standard exists because even expert neuroradiologists disagree on tumour boundaries — particularly at the oedema/infiltration margin on FLAIR and at the necrotic core boundary within the tumour.

Typical inter-reader variability in a well-run neuroradiologist annotation programme:

• Whole tumour: Mean Dice 0.88–0.93 between any two neuroradiologists
• Tumour core: Mean Dice 0.83–0.89 between any two neuroradiologists
• Enhancing tumour: Mean Dice 0.80–0.88 between any two neuroradiologists
• Necrotic core boundary: Mean Dice 0.71–0.82 between any two neuroradiologists

Cases with any pairwise Dice below 0.70 on any sub-region — typically 12–18% of a glioblastoma dataset — require adjudication by a senior neuroradiologist or neuro-oncologist to produce a consensus label. This adjudication step is where much of the annotation cost accumulates, but it is also where dataset quality is determined. Skipping adjudication and using the lower-quality of the two readers' annotations as the training label is a documented source of model performance ceiling.

Case Study: From Dice 0.67 to 0.94 on Glioblastoma Segmentation

A neuro-oncology AI developer had trained an initial tumour segmentation model on a dataset of 1,400 MRI cases annotated using general radiologists with a single-sequence (T1ce only) annotation protocol. Annotations were 2D contours on selected axial slices rather than 3D volumetric masks propagated across all slices.

Before: The model achieved a mean Dice coefficient of 0.67 on enhancing tumour and 0.71 on whole tumour on a 300-case external validation set curated from a separate clinical site. Error analysis revealed two systematic failure modes: (1) the model systematically under-segmented the necrotic core, which had not been separately annotated in training, and (2) the FLAIR oedema boundary was poorly learned because training annotations used only T1ce sequence data.

The reannotation programme involved:

• Reannotating all 1,400 cases with dual-reader four-sequence BraTS-protocol annotation by six board-certified neuroradiologists
• 3D volumetric masks for all three tumour sub-regions (ET, TC, WT) across all sequences
• Adjudication of the 218 cases (15.6%) with any pairwise sub-region Dice below 0.75, performed by a senior neuro-oncologist
• Full 21 CFR Part 11 compliant audit logs for all annotation events, annotator credentials, and adjudication decisions
• HIPAA-compliant data handling with BAA and de-identified DICOM throughout
• Expansion of the training set with 800 additional GBM cases and 400 lower-grade glioma cases to improve grade generalisation

After: The retrained model achieved a mean Dice of 0.94 on enhancing tumour, 0.92 on whole tumour, and 0.88 on tumour core on the same external validation set. Treatment response classification accuracy on a 120-case post-treatment dataset improved from 61% to 87%, with pseudo-progression versus true recurrence classification accuracy rising from 54% to 79%.

Total annotation cost for the reannotation and expansion programme was approximately AUD $1.4 million. The developer estimated this reduced time-to-regulatory-submission by 18 months compared to continued model architecture iteration on the original dataset.

FDA 21 CFR Part 11 and HIPAA: Compliance Requirements for Brain MRI Annotation

Brain MRI scans are protected health information (PHI) under HIPAA, and annotation workflows for FDA-submission neuro-oncology AI must satisfy both HIPAA security requirements and FDA 21 CFR Part 11 electronic records requirements. In practice, this means:

• HIPAA: Business Associate Agreement (BAA) between the clinical data holder and annotation vendor; all data transfers encrypted in transit (TLS 1.2+) and at rest (AES-256); de-identified DICOM headers before any data leaves the clinical institution; access restricted to credentialed annotators only; six-year documentation retention.
• 21 CFR Part 11: Immutable audit trail recording annotator identity (with credential documentation), annotation timestamps, modification history, adjudication decisions, and QC review events; electronic signature on final annotation records; controlled access with unique user IDs; no retroactive editing without logged justification.

Annotation vendors who cannot produce both a signed BAA and a 21 CFR Part 11 compliance attestation from their annotation platform provider are not appropriate for clinical AI submission datasets. The FDA has increasingly asked about annotation provenance in AI/ML SaMD submissions, and the absence of compliant audit trails is a material gap in regulatory submissions.

AI Taggers provides MRI annotation with full HIPAA-compliant data handling and 21 CFR Part 11 audit trail documentation for neuro-oncology AI datasets.

Annotation Cost Ranges for Brain Tumour MRI Datasets

Brain tumour MRI segmentation is among the most expensive annotation tasks in medical AI, driven by neuroradiologist credential requirements, multi-sequence complexity, and adjudication overhead. Realistic 2026 pricing for production-quality work:

• Single-reader, single-sequence, whole-tumour boundary only: AUD $120–$200 per MRI case
• Single-reader, four-sequence, three sub-regions (BraTS protocol): AUD $250–$420 per MRI case
• Dual-reader, four-sequence, three sub-regions with adjudication: AUD $500–$850 per MRI case
• Above with 21 CFR Part 11 provenance and HIPAA-compliant documentation: AUD $580–$1,000 per MRI case
• Longitudinal paired pre/post-treatment (per timepoint): AUD $350–$700 per scan per timepoint

At these rates, a 1,000-case dual-reader BraTS-protocol dataset with compliance documentation costs AUD $500,000–$1,000,000. This is not negotiable through lower-credential annotators — the FDA and IRBs reviewing AI/ML SaMD submissions are now routinely asking for annotator credential attestations and inter-reader agreement statistics as part of their review.

Related Medical Imaging Annotation Resources

Brain tumour MRI segmentation sits within the broader neuroimaging and medical AI annotation ecosystem. Related services and guides:

• MRI annotation services — full-service neuro, cardiac, and musculoskeletal MRI annotation
• Organ segmentation annotation — radiotherapy and surgical planning segmentation
• MRI annotation for neuroradiology AI: workflows that pass clinical review
• Organ segmentation annotation for surgical and radiotherapy AI
• FDA 21 CFR Part 11 for annotation provenance documentation

Frequently Asked Questions