Predicting brain tumour enhancement from non-contrast MR imaging with artificial intelligence: a multi-cohort retrospective diagnostic accuracy study

Technical Summary: Predicting Brain Tumour Enhancement from Non-Contrast MR Imaging with Artificial Intelligence

Problem Statement
Standard neuro-oncological imaging protocols typically require both pre- and post-contrast magnetic resonance imaging (MRI) to assess brain tumours. However, the administration of gadolinium-based contrast agents is not always feasible or desirable due to contraindications (e.g., renal impairment, allergy), patient refusal, or the desire to minimize cumulative exposure, particularly in paediatric populations and during frequent surveillance. Furthermore, gadolinium disposal poses environmental concerns. While deep learning has shown promise in reducing contrast dosage or synthesizing virtual contrast, existing evidence is largely confined to single-institution, adult, predominantly glioma cohorts of modest size, lacking large-scale, heterogeneous validation and direct benchmarking against expert radiologists on non-contrast tasks.

Methodology
The authors conducted a multi-cohort retrospective diagnostic accuracy study using a dataset of 11,089 brain MRI studies from 8,548 unique patients. The data was aggregated from 10 international datasets across four countries (North America, Europe, Africa) and three continents, spanning adult and paediatric populations with diverse pathologies (glioma, meningioma, metastases, and post-resection appearances). Imaging was acquired between 2006 and 2024 on 1.5T and 3T scanners from various manufacturers.

• Data Preprocessing: The study utilized only non-contrast sequences (T1-weighted, T2-weighted, and T2/FLAIR) as model inputs. Ground-truth labels for enhancing tumours were derived from complete datasets including post-contrast T1-weighted images. No site- or scanner-level intensity harmonization was applied to ensure the model learned from acquisition heterogeneity rather than being shielded from it.
• Model Development: Three fully volumetric multi-modal 3D deep learning architectures were trained: nnU-Net (v2), SegResNet, and SwinUNETR. A randomized train-test split at the patient level (90% training, 10% testing) was employed to prevent data leakage.
• Evaluation Metrics: Performance was assessed on a held-out test set of 1,109 studies. Primary endpoints included patient-level enhancement detection (balanced accuracy, sensitivity, specificity) and voxel-level segmentation quality (Dice coefficient). Clinical interpretability thresholds were defined as Dice ≥ 0.3 (acceptable), ≥ 0.5 (good), and ≥ 0.7 (excellent).
• Radiologist Benchmarking: Eleven expert radiologists reviewed a subset of 564 cases (100 cases each) from the test set under blinded conditions (no clinical history, prior imaging, or referral indication). They were tasked with predicting the presence of enhancement from non-contrast images alone.
• Equity Assessment: The study performed a comprehensive equity assessment across demographic (age, sex), clinical (pathology, post-operative status), and geographic subgroups.

Key Contributions

1. Scale and Heterogeneity: This is reported as the largest and most heterogeneous evaluation of contrast-enhancement prediction from non-contrast brain MRI to date, incorporating paediatric cohorts and diverse pathologies.
2. Direct Human-AI Benchmarking: The study provides the first direct comparison of deep learning performance against a blinded panel of expert radiologists on the specific task of predicting enhancement from non-contrast images.
3. Equity Analysis: A formal assessment of performance equity across demographic and geographic subgroups was conducted.
4. Failure Mode Analysis: The authors quantified drivers of detection failure, categorizing errors by lesion size, atypical enhancement patterns, and post-operative changes.

Results

• Model Performance: The nnU-Net architecture outperformed SegResNet and SwinUNETR. On the held-out test set, nnU-Net achieved a balanced accuracy of 83.0% (95% CI 79.1–87.2) for patient-level enhancement detection, with a sensitivity of 91.5% and specificity of 74.4%. For voxel-level segmentation, the mean Dice coefficient for enhancing tumours was 0.574 ± 0.319.
• Segmentation Quality: Among 992 enhancing cases, 76.8% achieved a Dice ≥ 0.3, 67.5% achieved Dice ≥ 0.5, and 50.2% achieved Dice ≥ 0.7. The model accurately predicted enhancement volume with an R² of 0.859.
• Comparison with Radiologists: The expert radiologists’ majority-vote performance was lower than the AI model, with a balanced accuracy of 71.7% (sensitivity 77.6%, specificity 65.8%). The model was significantly more accurate than the radiologist panel (McNemar’s test, p<0.0001).
• Subgroup Performance: Performance was generally robust across pathologies (e.g., meningioma 93%, presurgical glioma 76%, metastases 74%) but was notably lower in paediatric cases (45% reaching Dice ≥ 0.3). Performance did not differ significantly between men and women.
• Failure Analysis: Of the 260 cases failing to reach acceptable detection (Dice < 0.3), the primary causes were small lesion volume (<0.5 cm³, 40.4%), atypical enhancement patterns (38.5%), false positives (11.5%), and post-operative changes (6.5%).

Significance and Interpretation
The authors conclude that deep learning models can identify contrast-enhancing brain tumours from non-contrast MRI with higher balanced accuracy than expert radiologists performing the same unfamiliar task. While the current performance is insufficient to replace contrast-enhanced MRI outright—particularly in paediatric populations where the model is weaker—the findings support the use of such models as triage or decision-support adjuncts. Specifically, the model could flag studies where post-contrast enhancement is likely, prompting the addition of contrast to a non-contrast-only protocol, thereby avoiding second hospital visits and minimizing delays. Conversely, studies predicted to be enhancement-free might proceed without gadolinium with appropriate caution, potentially reducing gadolinium dependence in neuro-oncology imaging. The study emphasizes that future work should focus on optimizing models in collaboration with radiology experts and addressing performance disparities in underrepresented subgroups like paediatric patients.