Paper review · October 23, 2026
ViBMR: A Large-Scale Multi-Sequence Brain MRI Dataset and Comprehensive Evaluation for Vietnamese Radiology Report Generation
The authors built a multi-sequence brain MRI dataset paired with Vietnamese radiology reports from real hospitals and present report generation baselines by encoder, sequence and fusion strategy.
Problem
Data for brain MRI report generation is mostly in English. In many datasets, reports were also written later and attached to segmentation datasets, which differs from the actual reporting workflow in hospitals. No real clinical report data written in a low-resource language such as Vietnamese existed.
Method
The authors collected 3,649 patients and 11,875 MRI series from a tertiary hospital in Vietnam. The sequences are T1, T2, FLAIR and contrast-enhanced T1, and the findings section was extracted from de-identified reports as the ground truth. Volumes were resampled to 1mm isotropic resolution, and HD-BET was used to keep only the brain region. Per-sequence features were extracted with M2FTrans and mmFormer encoders pretrained on segmentation. These features were fused using averaging, learned weights, spatial and channel gating, cross-attention and other methods, then fed into an LLM decoder. The study has four research questions: the cross-dataset gap with MR-RATE, the effect of segmentation-pretrained encoders, a comparison of fusion strategies, and the impact of sequence combinations.
Results
- The most common sequence combination was T1, T2 and FLAIR, with 2,411 cases. 939 studies had all four sequences, and 283 had the T2, FLAIR and contrast-enhanced T1 combination.
- Findings followed a long-tail distribution. Small vessel disease was the most common, followed by sinus mucosal thickening, cerebral infarction and hemorrhage. A single report often contained multiple findings across multiple regions.
- Based on their experiments, the authors report that generating native Vietnamese reports reliably requires a real clinical dataset in that language. However, the provided text does not include specific numbers.
Limitations
Because the data comes from a single hospital, generalization to other institutions and scanners is limited. There are no segmentation annotations, and only the findings section of the reports is used as ground truth. The sequence composition is imbalanced, and rare findings have few samples. The dataset is planned to be released for non-commercial research use only. The provided text is cut off before the experimental results, so detailed numbers comparing model performance could not be included in this review.
Why it matters
It shows that medical AI in low-resource languages needs data based on real clinical reports, not translated or rewritten ones. It also provides baselines for how to fuse multi-sequence MRI.
Key terms
- Multi-sequence MRI
- Multiple images acquired under different imaging conditions, such as T1, T2, FLAIR and contrast-enhanced T1. They provide complementary information, such as edema, lesions and contrast enhancement.
- Radiology Report Generation
- The task of taking a medical image as input and automatically producing the findings text that a radiologist would write.
- Modality fusion
- The process of combining features extracted from each sequence into a single representation, using methods such as averaging, weighting, gating or attention.
Original
Authors: Dai Lam Le, Minh Khoi Nguyen, Son Dinh Nguyen, Amir Reza Jafari, Viet Ha, Nguyen Son, Reza Farahbakhsh, Noel Crespi, et al. · Posted 2026-10-21 · License arXiv 비독점 배포 라이선스 · Based on the full text
This review is an AI-written summary of the paper. Please treat the original paper as the reference.