ViBMR: A Large-Scale Multi-Sequence Brain MRI Dataset and Comprehensive Evaluation for Vietnamese Radiology Report Generation
Authors: Dai Lam Le, Minh Khoi Nguyen, Son Dinh Nguyen, Amir Reza Jafari, Viet Ha, Nguyen Son, Reza Farahbakhsh, Noel Crespi, Phi Le Nguyen
Organizations: AI4LIFE, Hanoi University of Science and Technology, Vietnam · SAMOVAR, Télécom SudParis, Institut Polytechnique de Paris, France · Concordia International School Hanoi, Vietnam · Hanoi – Amsterdam High School for the Gifted, Vietnam
Automated radiology report generation for multi-sequence brain MRI holds strong promise for alleviating radiologists' workload, yet it remains a demanding problem: brain MRI studies are inherently volumetric, each pulse sequence carries complementary diagnostic information, and annotated data are critically scarce for low-resource clinical languages. Existing brain MRI-report resources are still dominated by English reports, translated clinical text, or datasets adapted from task-specific imaging benchmarks, leaving routine low-resource clinical reporting and heterogeneous multi-sequence MRI protocols insufficiently represented. To bridge this gap, we introduce ViBMR, to the best of our knowledge, the first large-scale multi-sequence brain MRI dataset paired with routine Vietnamese clinical reports, comprising 3,649 patients and 11,875 MRI series across T1, T2, FLAIR, and contrast-enhanced T1 sequences. Building on ViBMR, we establish a comprehensive evaluation setup for brain MRI report generation, covering representative medical vision-language models, multi-sequence inputs, and feature aggregation strategies. The results provide key insights into dataset characteristics and model behavior, while demonstrating the necessity of native low-resource clinical datasets for reliable Vietnamese-origin brain MRI report generation. Our dataset and baseline models will be released for non-commercial research.
Figures & tables
Figure 1: Overview of the proposed brain MRI report generation framework. A multi-sequence MRI encoder is pretrained on segmentation and its weights are reused to extract visual tokens from FLAIR, T1c, T1, and T2 sequences. The fused visual tokens are projected and concatenated with instruction prompt tokens, then fed into an LLM decoder to generate the final MRI report.
Dataset
Anatomy
MRI sequences
Seg.
Clin.
Scale
MR-RATE [ 10 ]
Brain, Spine
3D volumetric MRI
✗
✓
83,425 patients / 705,254 volumes
SSPH Dataset [ 24 ]
Brain
T1, T2, FLAIR, DWI
✓
✗
12,535 cases / 30,793 pairs
BTReport-BraTS [ 14 ]
Brain
T1, T1c, T2, FLAIR
✓
✗
1,470 cases
RadGenome -Brain MRI [ 24 ]
Brain
T1, T1c, T2, FLAIR, DWI, ADC
✓
✗
1,007 cases / 3,408 pairs
HuskyBrain [ 14 ]
Brain
T1,T1c, T2, FLAIR
✓
✗
184 patients
ViBMR (Ours)
Brain
T1,T1c, T2, FLAIR
✗
✓
3,649 patients / 11,875 series
Table 1: Comparison with representative MRI datasets related to report generation and 3D MRI understanding. “Seg.” indicates whether segmentation annotations are provided. “Clin.” indicates whether the dataset provides full-study routine clinical reports directly collected from the clinical workflow.
Figure 2: Dataset and distribution analysis of the MRI report dataset. (a) Distribution of MRI sequence availability across the dataset. The bar values above each modality indicate the number of cases containing that sequence, while connected black dots represent specific modality combinations and their corresponding case counts among four MRI sequences: T1-weighted imaging (T1), T2-weighted imaging (T2), Fluid-Attenuated Inversion Recovery (FLAIR), and contrast-enhanced T1-weighted imaging (T1c); (b) Top 10 most frequent abnormal findings; (c) Top 10 most frequently mentioned anatomical regions, comparing their prevalence among all cases and abnormal cases.
Figure 3: Overview of the proposed MRI report generation pipeline. Raw MRI slices from axial, sagittal, and coronal views are first reconstructed into 3D volumes with different input shapes, followed by voxel-size normalization, brain cropping, and organization into four cropped 3D MRI modalities: FLAIR, T1c, T1, and T2. Each modality is processed by a separate encoder, and the resulting modality-specific visual features are fused before token extraction. The zoom-in panel illustrates the fusion variants evaluated in this study, including mean fusion, learnable weighted fusion, spatial gating, channel gating, mean-query cross-attention, and the M2FTrans-MMFT bottleneck-based fusion strategy.
Figure 4: Cross-dataset generalization after language normalization. The model is trained on MR-RATE and evaluated on both MR-RATE and ViBMR. Decrease denotes the performance decrease from MR-RATE to ViBMR. A larger Decrease indicates a larger generalization gap.
M2FTrans (Encoder)
mmFormer (Encoder)
AutoRG-Brain
M3D
BLEU-1
68.09
64.31
54.44
42.09
BLEU-4
60.04
58.22
47.25
35.32
ROUGE-1
75.02
72.59
78.12
57.06
ROUGE-2
66.67
65.68
62.80
47.52
ROUGE-L
70.17
68.69
65.85
50.94
BERT
87.02
85.55
83.30
80.80
Table 2: Comparison of brain MRI report generation model performance under the same LoRA adapter setting for decoder models. M2FTrans and mmFormer are used as encoder baselines with Qwen3-4B-Instruct as the language model. BERT denotes the BERTscore F1, while CLIN denotes the Clinical-F1 score. The best value for each metric is shown in bold , and the second-best value is underlined .
Encoder
Fusion
BLEU-4
ROUGE-2
BERT
CLIN
M2FTrans
mean
59.13
72.62
87.25
72.33
weighted
51.02
66.75
85.48
68.56
spatial_gate
58.32
72.24
87.25
71.62
channel_gate
58.95
72.64
87.37
70.41
crossattn_meanq
53.8
68.95
85.89
69.84
m2ftrans_mmft
58.3
72.22
87.10
70.84
Table 3: Comparison of different fusion strategies for MRI report generation using visual encoders initialized from M2FTrans and mmFormer checkpoints. The best value in each encoder group is shown in bold , and the second-best value is underlined . The final row in each group denotes the original baseline configuration.
Encoder
Method
Modality
B-4
R-L
BERT
CLIN
M2FTrans
mean
T1
40.70
58.99
80.25
50.09
T2
40.94
61.46
82.57
62.79
FLAIR
48.19
66.62
83.66
61.83
T1+T2
50.53
67.52
84.43
62.57
T1+T2+FLAIR
59.29
74.55
87.21
71.75
All
53.77
70.58
85.15
69.71
Table 4: Comparison of modality importance across different fusion methods using M2FTrans and mmFormer encoder checkpoints. In the Modality column, the ”+” symbol denotes combining additional sequences. ”All” denotes using all four available MRI modalities. B-4 and R-L denote BLEU-4 and ROUGE-L, respectively. The best value for each metric within each method is shown in bold , and the second-best value is underlined .
University of Naples Federico II, Department of Electrical Engineering and Information Technology, Italy · Northwestern University, Dept. of Computer Science, McCormick School of Engineering and Applied Science, United States · Pegaso University, Department of Information Science and Technology, Italy
Department of Bioengineering, Imperial College London, London, UK · Imperial-X, Imperial College London, London, UK · School of Biomedical Engineering & Imaging Sciences, King’s College London, London, UK