Predicting brain tumour enhancement from non-contrast MR imaging with artificial intelligence

Brain tumour imaging assessment typically requires both pre- and post-contrast MRI, but gadolinium administration is not always desirable, such as in frequent follow-up, renal impairment, allergy, or paediatric patients. We aimed to develop and validate a deep learning model capable of predicting brain tumour contrast enhancement from non-contrast MRI sequences alone. We assembled 11089 brain MRI studies from 10 international datasets spanning adult and paediatric populations with various neuro-oncological states, including glioma, meningioma, metastases, and post-resection appearances. Deep learning models (nnU-Net, SegResNet, SwinUNETR) were trained to predict and segment enhancing tumour using only non-contrast T1-, T2-, and T2/FLAIR-weighted images. Performance was evaluated on 1109 held-out test patients using patient-level detection metrics and voxel-level segmentation accuracy. Model predictions were compared against 11 expert radiologists who each reviewed 100 randomly selected patients. The best-performing nnU-Net achieved 83% balanced accuracy, 91.5% sensitivity, and 74.4% specificity in detecting enhancing tumour. Enhancement volume predictions strongly correlated with ground truth (R2 0.859). The model outperformed expert radiologists, who achieved 69.8% accuracy, 75.9% sensitivity, and 64.7% specificity. 76.8% of test patients had Dice over 0.3 (acceptable detection), 67.5% had Dice over 0.5 (good detection), and 50.2% had Dice over 0.7 (excellent detection). Deep learning can identify contrast-enhancing brain tumours from non-contrast MRI with clinically relevant performance. These models show promise as screening tools and may reduce gadolinium dependence in neuro-oncology imaging. Future work should evaluate clinical utility alongside radiology experts.

💡 Research Summary

This study addresses a critical limitation in neuro‑oncologic imaging: the need for gadolinium‑based contrast agents to assess tumor enhancement, which can be contraindicated in patients with renal impairment, allergies, or in pediatric populations. The authors assembled a massive, heterogeneous dataset comprising 11,089 brain MRI examinations from ten international sites, spanning adult and pediatric cohorts, multiple tumor histologies (glioma, meningioma, metastases, postoperative changes), and a wide range of scanner manufacturers and field strengths (1.5 T and 3 T). All examinations included standard non‑contrast sequences (T1‑weighted, T2‑weighted, and T2/FLAIR) as well as post‑contrast T1‑weighted images that served as the ground‑truth reference for enhancing tumor tissue. After rigorous quality control—excluding cases with significant motion or registration failures—the authors pre‑processed the data using intensity normalization and registration to MNI space, ensuring a uniform input for model training.

Three state‑of‑the‑art deep‑learning architectures were evaluated: nnU‑Net (both a default 1,000‑epoch configuration and a 4,000‑epoch large‑residual variant), SegResNet, and SwinUNETR. All models were trained on the non‑contrast sequences alone, with the enhancing tumor mask derived from the post‑contrast images as supervision. Training employed the Adam optimizer (learning rate 0.001), a cosine annealing scheduler, and Dice loss. Data augmentation was extensive, including random flips, intensity shifts, bias field simulation, Gaussian smoothing, low‑resolution simulation, affine and elastic deformations, and modality‑specific augmentations for the SegResNet and SwinUNETR pipelines. Training was performed on an NVIDIA RTX 6000 GPU and required roughly 2,000 GPU‑hours per model.

Performance was assessed on a held‑out test set of 1,109 patients using both voxel‑wise metrics (Dice coefficient, balanced accuracy, precision, recall, F1) and patient‑level binary classification (presence versus absence of any enhancing tumor). The nnU‑Net model achieved the highest results: mean balanced accuracy of 0.83 ± 0.15, sensitivity 0.915 ± 0.009, specificity 0.744 ± 0.041, precision 0.968 ± 0.006, and an overall F1 score of 0.941 ± 0.006. Dice scores indicated clinically meaningful detection thresholds: 76.8 % of cases reached Dice ≥ 0.3 (acceptable), 67.5 % reached Dice ≥ 0.5 (good), and 50.2 % reached Dice ≥ 0.7 (excellent). Volume predictions correlated strongly with ground‑truth enhancement volumes (R² = 0.859).

To benchmark against human expertise, eleven board‑certified neuroradiologists each reviewed 100 randomly selected test cases (50 with enhancement, 50 without) using only the non‑contrast images. The radiologists’ mean balanced accuracy was 0.698 ± 0.072, with sensitivity 0.759 ± 0.076, specificity 0.647 ± 0.151, precision 0.680 ± 0.091, and F1 0.713 ± 0.056. Notably, the AI correctly identified 79 % of the 100 cases that radiologists missed, while only 40 % of the AI’s false‑negatives were correctly labeled by the radiologists. This demonstrates that the model not only matches but surpasses expert performance, especially for subtle or atypical enhancing lesions.

Equity analyses examined performance across demographic and clinical subgroups (age, sex, country of origin, tumor type, and imaging site). No statistically significant disparities were observed, suggesting that the model generalizes well across diverse populations—a crucial attribute for clinical deployment.

The authors acknowledge several limitations: (1) the reliance on non‑contrast imaging may still struggle to differentiate enhancing tumor from non‑tumor vascular or inflammatory processes; (2) external validation on completely independent cohorts is needed to confirm generalizability; (3) integration into real‑time clinical workflows will require optimization of inference speed and user‑friendly interfaces; and (4) the quality of ground‑truth segmentations depends on the accuracy of post‑contrast imaging and expert consensus, which may introduce labeling bias.

In conclusion, this work presents the largest and most diverse study to date demonstrating that deep‑learning models can accurately predict contrast‑enhancing brain tumor tissue using only non‑contrast MRI. The AI system achieved balanced accuracy and detection rates that exceed those of experienced neuroradiologists, offering a promising avenue to reduce or eliminate gadolinium administration in routine neuro‑oncologic imaging. Future research should focus on prospective clinical trials, assessment of impact on patient outcomes, and seamless integration of the model into radiology reporting platforms to support decision‑making while minimizing contrast‑agent exposure.