BTReport-BraTS: Radiology Report Dataset
- BTReport-BraTS is an open‐source dataset that pairs BraTS’23 glioma mpMRI imaging with deterministic feature extraction for radiology report generation.
- It integrates anatomical descriptors, synthetic reports, structured metadata, and midline segmentations to support neuro‐oncology research.
- The framework employs a two-stage process—deterministic image feature extraction followed by LLM-based narrative synthesis—to produce auditable and clinically grounded reports.
BTReport-BraTS is an open-source companion dataset for brain tumor radiology report generation that augments BraTS'23 adult glioma mpMRI cases with anatomical descriptors, extracted metadata, structured summary reports, synthetic radiology reports generated by BTReport, and midline segmentations. It was built from the combined training and validation splits of BraTS 2023 Adult Glioma, totaling n = 1,470 pre-operative mpMRI cases. In the associated BTReport framework, radiology report generation is decomposed into deterministic feature extraction and LLM-based report composition, with the stated goal of producing reports that are completely interpretable and less prone to hallucinations than end-to-end vision-language approaches (Rivera et al., 17 Feb 2026).
1. BraTS context and problem setting
BraTS has, since 2012, evaluated state-of-the-art machine learning methods for glioma segmentation, classification, and survival prediction, and successive BraTS extensions have broadened the benchmark family to underrepresented MRI populations and new modalities. BraTS-Africa introduced Sub-Saharan Africa clinical MRI as a benchmark domain; BraTS-Path introduced histopathologic patch classification from H&E-stained, FFPE tissue sections; and BraTS-PEDs established the first BraTS challenge focused on pediatric brain tumors (Adewole et al., 2023, Bakas et al., 2024, Kazerooni et al., 2024).
Within that ecosystem, BTReport-BraTS addresses a different bottleneck. The central motivation is that neuro-oncology report generation lacks open paired image-text data, whereas conventional BraTS releases historically provide imaging and segmentation labels rather than paired radiology reports. BTReport-BraTS fills that gap by pairing BraTS adult glioma imaging with synthetically generated reports produced from clinically grounded, deterministically extracted features rather than direct free-form image captioning (Rivera et al., 17 Feb 2026).
This positioning is methodologically important. BraTS datasets have primarily supported segmentation, subregion delineation, lesion-wise evaluation, and challenge-style benchmarking, whereas BTReport-BraTS is intended to support brain tumor radiology report generation, neuro-oncology multimodal research, training / benchmarking report generation models, clinical-quality evaluation of synthetic reports, and survival and outcome modeling from imaging-derived features (Rivera et al., 17 Feb 2026).
2. Dataset composition and data products
BTReport-BraTS is derived from BraTS'23 adult glioma pre-operative mpMRI. The framework explicitly operates on the standard structural sequences T1n / T1, T1c / T1-Gd, T2w, and T2-FLAIR. The report-generation prompt explicitly instructs the LLM not to mention diffusion, perfusion, spectroscopy, MRA, or other modalities unless provided in metadata, thereby constraining the report content to available inputs (Rivera et al., 17 Feb 2026).
The dataset products are not limited to generated text. BTReport-BraTS includes imaging-linked metadata and intermediate representations intended to make the report-generation process inspectable. The package contains anatomical descriptors, extracted metadata, structured summary reports, synthetic radiology reports, and midline segmentations. A separate HuskyBrain dataset of 184 retrospective GBM cases from UWMC, with pre-operative mpMRI, tumor masks, and radiologist-authored reports, is used for report-quality evaluation against real clinical reporting. A separate BraTS'23-derived survival subset of n = 461 includes age at diagnosis, sex, MGMT methylation, IDH1 mutation status, and overall survival / days to known death (Rivera et al., 17 Feb 2026).
A key practical implication is that BTReport-BraTS is not merely a text augmentation of BraTS. It is a structured image-report resource in which the generated narrative is coupled to explicit imaging features and segmentations. This suggests its intended role is both as a report corpus and as a scaffold for measurement-grounded neuro-oncology AI.
3. Deterministic imaging feature extraction
BTReport is a two-stage framework whose first stage extracts clinically relevant features algorithmically from mpMRI and segmentation outputs. The anatomical pipeline is designed to be robust to tumor deformation: SynthMorph registers the MNI152 atlas into subject space, the warped atlas is treated as a pseudo-healthy anatomy, and SynthSeg is then run on this pseudo-healthy image to obtain anatomical labels. These labels are merged with tumor and midline segmentations to compute tumor volume, ventricle volume, lesion count, subregion proportions, tumor-ROI overlap, and tumor laterality/location descriptors (Rivera et al., 17 Feb 2026).
The framework also proposes a novel atlas-based 3D midline shift algorithm. A hand-annotated midline from an MNI152 atlas is registered into patient space using SynthMorph and compared with an “ideal” midline defined by connecting the anterior and posterior points of the falx cerebri in each axial slice. Voxel-wise distances between these lines provide a 3D MLS estimate. In parallel, BTReport uses a modified VASARI-auto pipeline. The paper characterizes VASARI as Visually AcceSAble Rembrandt Images and notes prior work showing that VASARI features can predict tumor grade, disease progression, IDH1 mutation status, MGMT methylation, recurrence risk, and overall survival (Rivera et al., 17 Feb 2026).
The paper lists 22 BTReport features in three groups:
| Group | Features |
|---|---|
| Segmentation statistics | Total tumor volume (mL); 3D lesion sizes (cm); Proportion of necrosis; Number of lesions; Proportion of enhancing tumor; Proportion of edema |
| VASARI-derived features | Ventricular invasion; Side of tumor epicenter; Enhancement quality; Enhancement thickness; Multiple satellites present; Multifocal or multicentric; Cortical involvement; Deep white matter invasion; Tumor location; Eloquent brain involved |
| Midline / mass effect features | Level of max MLS; Max MLS (mm) + left/right direction; Edema crosses midline; Enhancing tumor crosses midline; Asymmetrical ventricles; Enlarged ventricles |
These features are also the bridge between image analysis and narrative text. The report prompt maps metadata to findings about mass effect / ventricles and brain / enhancement, including direction and magnitude of midline shift, ventricular effacement or asymmetry, lesion location and laterality, tumor dimensions in three orthogonal axes, multifocality, enhancement pattern and thickness, necrosis proportion, edema extent, and cortical, deep white matter, or ependymal invasion (Rivera et al., 17 Feb 2026).
4. Report synthesis and linguistic control
The second BTReport stage uses a general-purpose LLM for narrative synthesis only. The paper experiments with gpt-oss:120B and LLaMA 3.1 70B Instruct, run locally/offline to avoid cloud-sharing of medical data. The LLM is given example FINDINGS sections from real brain tumor reports, extracted metadata, and stylistic instructions to mimic institutional radiology phrasing (Rivera et al., 17 Feb 2026).
The prompting strategy is restrictive by design. The model is instructed to use only metadata-supported facts, generate 10–20 clinically meaningful findings, preserve subsection structure, prioritize abnormal findings, report measurements in clinically conventional language, and avoid unsupported modalities or unsupported features. The framework therefore does not use the LLM for image interpretation; it uses the LLM for syntactic structuring and narrative formatting after deterministic image analysis has already been completed (Rivera et al., 17 Feb 2026).
This architecture differentiates BTReport-BraTS from direct image-to-text generation. The paper explicitly frames the system as avoiding end-to-end vision-language hallucination by grounding report generation in reliably computed quantitative imaging features. A plausible implication is that report interpretability is meant to arise from traceability: each report statement is intended to be attributable to a specific extracted feature or metadata field.
5. Clinical validation and reported performance
The clinical relevance of BTReport features is evaluated in two ways. First, using HuskyBrain reports, factual claims are extracted with TBFact, embedded with all-MiniLM-L6-v2, clustered with hierarchical agglomerative clustering using cosine distance and average linkage, and summarized using Gemma 3 27B. The resulting common report concepts include lateral ventricle effacement/asymmetry, midline shift, edema, lesion size, mass effect, ventricular enlargement, herniation, and corpus callosum involvement. The paper reports that 21/22 BTReport features are commonly reported in real radiology reports and 11/22 are statistically significant predictors of overall survival (Rivera et al., 17 Feb 2026).
Second, report generation quality is evaluated on 129 paired GBM image-report cases from HuskyBrain. For feature extraction reliability, the paper compares BTReport-extracted features against feature extractions from ground-truth reports using LangExtract. Reported categorical accuracy includes Side of tumor epicenter: 1.00 for both variants, Cortical involvement: about 0.79–0.80, and Ventricular effacement: about 0.52–0.56. Reported mean absolute errors include Midline shift: about 1.21–1.26 mm, Number of lesions: about 0.41–1.48, and Tumor volume: about 1.80–2.00 cm³. The paper states that improved segmentation input (V2: “Faking It” / nnU-Net-based winning submission) generally reduced error versus DeepMedic (V1) (Rivera et al., 17 Feb 2026).
For report generation, BTReport is compared with AutoRG-Brain and Seg-to-Exp:
| System | Lexical similarity | Factual alignment |
|---|---|---|
| BTReport (best variant) | BLEU-1 up to 0.248; BLEU-2 up to 0.136; ROUGE-1 up to 0.371; ROUGE-2 up to 0.115 | TBFact F1 0.359; BERTScore 0.453; RaTEScore 0.577 |
| AutoRG-Brain | BLEU-1 0.158; ROUGE-1 0.268 | TBFact F1 0.196; BERTScore 0.327; RaTEScore 0.477 |
| Seg-to-Exp | BLEU-1 0.085; ROUGE-1 0.163 | TBFact F1 0.098; BERTScore 0.156; RaTEScore 0.409 |
The paper states that all BTReport variants were significantly better than the baselines with p < 0.0001 and describes BTReport-generated reports as more closely aligned with reference clinical reports, more factually grounded, less hallucination-prone, better preserving clinical structure, and interpretable because every statement can be traced to extracted metadata (Rivera et al., 17 Feb 2026).
6. Relation to report-supervised MRI research and future directions
BTReport-BraTS belongs to a broader BraTS-adjacent movement toward report-conditioned brain MRI modeling. In longitudinal glioma change detection, report-generated weak labels increased dataset size by more than 3× and improved VGG classification from 75% to 82% AUC, with mixed training from scratch outperforming fine-tuning and feature extraction (Noto et al., 2022). In report-supervised segmentation, MS-RSuper used global quantitative and modality-wise qualitative findings on 1238 report-labeled BraTS-MEN/MET scans and outperformed both a sparsely supervised baseline and a naive report-supervision method by using substructure-aware, one-sided, and uncertainty-aware constraints (Ge et al., 24 Feb 2026). In report generation proper, Brain3D later argued for native 3D neuroradiology report generation and reported Clinical Pathology F1 of 0.951 versus 0.413 for a strong 2D baseline while maintaining perfect specificity on healthy scans (Barone et al., 25 Feb 2026). Large-scale report-aligned pretraining has also expanded through brat, which used approximately 77,228 brain MRI–report pairs and evaluated transfer to BraTS segmentation and report generation tasks (Kayser et al., 21 Dec 2025).
The BTReport paper is also explicit about limitations. The current prompt handles only a limited set of imaging features, excludes modalities not available in the base structural MRI sequences, and proposes future incorporation of white matter hyperintensities, basal cistern status, additional mpMRI modalities, and ischemic/hemorrhagic stroke findings. The paper also introduces BTReview, a web-based radiologist evaluation platform for synthetic report assessment, which suggests that formal human validation remains an active requirement rather than a completed component of the framework (Rivera et al., 17 Feb 2026).
Taken together, BTReport-BraTS is best understood as a BraTS-derived image-report resource and associated methodology for interpretable, measurement-grounded neuro-oncology reporting. Its distinctive contribution is not merely synthetic text generation, but the coupling of BraTS imaging to deterministic anatomical, tumor, VASARI, and midline-shift features that can be rendered into radiology-style prose while remaining explicitly auditable (Rivera et al., 17 Feb 2026).
