BTReport: Deterministic Brain Tumor Report Generator

• BTReport is an open-source two-stage framework that deterministically extracts clinically validated MRI features for brain tumor report generation.
• It separates image interpretation from report composition by using rule-based feature extraction and a large language model for narrative formatting.
• The system is clinically aligned, enhancing reproducibility and auditability while reducing hallucination risk in radiology reports.

BTReport is an open-source two-stage framework for brain tumor radiology report generation that constructs natural-language radiology reports from deterministically extracted imaging features rather than from an end-to-end vision–LLM. Its central design principle is a strict separation between image interpretation and report composition: deterministic, clinically motivated features encode what the MRI shows, while a LLM is used only for syntactic structuring and narrative formatting. In the 2026 formulation, this design is presented as a response to the absence of open paired MRI–report datasets in neuro-oncology, the quantitative and spatial complexity of brain tumor reporting, and the need for interpretable, auditable, and reproducible report generation; the same work also introduces the companion BTReport-BraTS dataset (Rivera et al., 17 Feb 2026).

1. Conceptual basis and problem setting

BTReport is situated within radiology report generation (RRG), but it departs from the dominant end-to-end paradigm that maps images directly to text. The framework is motivated by three conditions emphasized for neuro-oncology: large open tumor MRI datasets exist, but there are essentially no open paired MRI–report datasets; brain tumor reports depend on detailed anatomy, quantitative measurements, mass-effect descriptors, and standardized vocabularies such as VASARI; and generic vision–LLMs can hallucinate findings or fail on fine-grained spatial reasoning (Rivera et al., 17 Feb 2026).

Within this setting, BTReport defines a more modular interface between vision and language. Stage 1 performs deterministic extraction of clinically validated imaging features such as volumes, locations, midline shift, and VASARI labels. Stage 2 conditions a LLM on those extracted features and constrains it to produce only a report narrative grounded in the metadata. The paper states three explicit aims for this design: to use deterministic, interpretable, clinically validated imaging features as the interface between vision and language; to use LLMs only for stylistic and narrative synthesis, not for measurement or primary image interpretation; and to make the entire pipeline transparent, reproducible, and auditable (Rivera et al., 17 Feb 2026).

This division of labor distinguishes BTReport from methods that entangle visual inference and text generation in a single neural system. A practical implication is that each report can be decomposed into an explicit feature representation and a formatting layer. The paper treats this as a means of reducing hallucination risk and improving traceability, because any incorrect report can in principle be attributed either to feature extraction error or to prompt-following failure at the language stage (Rivera et al., 17 Feb 2026).

2. System architecture and processing pipeline

The BTReport pipeline begins with structural MRI and tumor segmentation, then moves through deterministic feature extraction, structured metadata assembly, and constrained report generation. The high-level inputs are T1-weighted MRI and, for tumor segmentation, the full mpMRI set. Tumor masks may be manual or automatic. The resulting feature set includes tumor-robust anatomical segmentations, volumetric statistics, 3D midline shift measurements, and VASARI features, all of which are collected into a structured metadata representation supplied to the LLM (Rivera et al., 17 Feb 2026).

Preprocessing follows BraTS/CaPTk conventions. The paper specifies DICOM-to-NIfTI conversion, registration to SRI24 or MNI-like space with 1 mm isotropic resampling, and skull stripping via CaPTk. For HuskyBrain, tumor segmentations are obtained either from DeepMedic or from the BraTS 2023 winning nnU-Net ensemble, “Faking It.” This establishes a normalized geometric frame in which deterministic spatial measurements and atlas-based analyses can be performed consistently across cases (Rivera et al., 17 Feb 2026).

A key architectural choice is the use of tumor-robust anatomical segmentation rather than direct anatomical inference on tumor-bearing scans. Standard MNI152 atlas space is first registered to the subject T1 with SynthMorph, producing an MNI152-to-subject transformation. The warped atlas serves as a pseudo-healthy anatomy, after which SynthSeg is run on the MNI152-to-subject volume. The paper argues that because the input is pseudo-healthy, anatomical boundaries remain robust even in the presence of large tumors. These labels are then merged with tumor subregions and midline segmentation to form a unified label map spanning normal structures, pathological subregions, and mass-effect structures (Rivera et al., 17 Feb 2026).

The generated reports currently target only the Findings section. This delimitation is important: BTReport is not presented as a full report authoring system for impression or recommendation generation, but as a structured neuro-oncology findings generator grounded in deterministic imaging features (Rivera et al., 17 Feb 2026).

3. Deterministic feature extraction and image-derived representation

BTReport’s deterministic representation is the core of the framework. From the unified label map, it computes total tumor volume in milliliters as the sum of tumor subregion voxels times voxel volume, together with subregion proportions for necrosis, enhancement, and edema. It also derives lesion-level features, including 3D lesion size in centimeters via axis-aligned bounding-box extents in the anteroposterior, transverse, and craniocaudal dimensions, the number of lesions, and multifocal versus multicentric status based on connectivity and spatial separation (Rivera et al., 17 Feb 2026).

The framework also computes tumor–ROI relations through overlap with SynthSeg labels and atlas-defined masks. These include tumor location in lobes and deep structures, cortical involvement, deep white matter invasion, ventricular invasion through subependymal or ependymal overlap, and eloquent brain involvement, including motor, language, and visual cortex masks. The paper groups these features into segmentation statistics, VASARI-derived and anatomical descriptors, and midline or mass-effect descriptors (Rivera et al., 17 Feb 2026).

A distinct contribution is the 3D midline shift pipeline. In MNI152 space, the authors manually annotate midline structures including the falx cerebri, septum pellucidum, and ventricles, together with an ideal midline axis formed by connecting anterior and posterior falx cerebri points in each axial slice. SynthMorph is then used to warp this template into subject space. For each axial slice, BTReport computes signed left or right deviations in millimeters between subject-specific midline structures and the ideal midline axis. From these measurements it derives maximum MLS and direction, the axial level of maximum shift, and binary descriptors including whether edema or enhancing tumor crosses midline, whether ventricles are asymmetrical, and whether ventricles are enlarged (Rivera et al., 17 Feb 2026).

VASARI extraction is incorporated through a modified variant of VASARI-auto that consumes subject-specific SynthSeg anatomy and midline segmentations. The resulting descriptors include tumor location, enhancement quality, enhancement thickness, cortical involvement, deep white matter invasion, ventricular or ependymal invasion, multifocality or multicentricity, satellite lesions, and eloquent brain involvement. The framework is described as deterministic in the sense that, given a fixed MRI, segmentation model, and registration algorithm, the resulting features are fixed. Because the tools are open source and the features are explicit numerical or categorical quantities, the authors present the representation as reproducible and directly auditable (Rivera et al., 17 Feb 2026).

Feature family	Representative outputs	Source in pipeline
Segmentation statistics	Total tumor volume, necrosis/enhancement/edema proportions, lesion size, lesion count	Tumor masks and voxel geometry
Anatomical and VASARI descriptors	Tumor location, cortical involvement, deep white matter invasion, ventricular invasion, eloquent brain involvement	SynthSeg anatomy, atlas overlaps, modified VASARI-auto
Midline and mass-effect descriptors	Maximum MLS, shift direction, level of maximum MLS, edema or enhancing tumor crossing midline, ventricular asymmetry	Atlas-based 3D MLS pipeline

The paper further validates the feature space along two axes. First, 21 of 22 BTReport features correspond to one of the top 35 radiology topics mined from real reports. Second, 11 of 22 are statistically significant predictors of overall survival in BraTS’23. This suggests that the representation is not merely convenient for text generation, but also clinically aligned and prognostically informative (Rivera et al., 17 Feb 2026).

4. Language-model report generation and grounding constraints

The report-generation stage uses LLMs only after all clinically relevant content has been converted into structured metadata. BTReport evaluates two open-source LLMs, gpt-oss:120B and LLaMA 3.1 70B Instruct, and the paper states that neither is fine-tuned on radiology reports. Inference is performed locally and offline so that patient data are not shared with external services (Rivera et al., 17 Feb 2026).

Prompt construction is central to the framework. The prompt includes a radiologist role specification, multiple example Findings sections from real HuskyBrain reports for in-context stylistic guidance, the metadata JSON for the subject, and explicit global rules constraining the permissible content. Among the quoted rules are: use only the metadata provided for quantitative statements; do not hallucinate information that is not directly inferable; preserve subsection structure from the example reports; include the subsections MASS EFFECT & VENTRICLES and BRAIN / ENHANCEMENT; and never mention imaging sequences other than T1n, T2w, T2 FLAIR, or T1-Gd unless stated explicitly in metadata (Rivera et al., 17 Feb 2026).

The prompt also encodes mandatory clinical considerations. These include explicit reporting of maximum midline shift in millimeters with direction and anatomical level, mass effect on ventricles and surrounding structures, ventricular effacement with side and horn, 3D tumor measurements, lesion number, dominant lesion and laterality, enhancement characteristics and proportion, edema volume and whether edema crosses midline, cortical and deep white matter invasion, ventricular invasion, eloquent brain involvement, and the proportion of necrosis for describing central necrosis. The short prompt version retains the same control philosophy while emphasizing top metadata-supported findings and prohibiting unsupported modalities such as diffusion, perfusion, spectroscopy, or MRA (Rivera et al., 17 Feb 2026).

Grounding is therefore achieved not by latent multimodal alignment inside a single model, but by prompt-level confinement to a deterministic metadata interface. The metadata JSON functions as the sole structured source of report content, while the LLM determines ordering, phrasing, and stylistic conformity. The paper reports that this design yields substantially higher factual inclusion and lower distortion or omission than the evaluated baselines, which the authors interpret as evidence that the constrained prompting regime is effective at hallucination mitigation (Rivera et al., 17 Feb 2026).

5. Clinical relevance, prognostic value, and report-quality evaluation

BTReport evaluates its feature representation before evaluating report text. To test alignment with real radiology discourse, the authors extract factual claims from HuskyBrain reports with TBFact using DeepSeek-R1, embed those claims with sentence-transformers all-MiniLM-L6-v2, cluster them by hierarchical agglomerative clustering with cosine distance and average linkage, and summarize clusters with Gemma 3 27B. The top 35 clusters cover concepts such as lesion size, midline shift magnitude, ventricular effacement or asymmetry, edema, herniation, and corpus callosum involvement. BTReport then maps its features to these topics and reports that 21 of 22 align with a top-35 concept (Rivera et al., 17 Feb 2026).

The paper also analyzes prognostic value on a subset of BraTS’23 with clinical metadata, comprising $n=461$ cases with age, sex, MGMT methylation, IDH1 status, and overall survival time. Kaplan–Meier curves, log-rank tests, and Cox proportional hazards models are used, with the standard form

$h(t \mid x) = h_0(t)\exp(\beta^T x).$

The reported outcome is that 11 of 22 BTReport features are statistically significant predictors of overall survival. Examples explicitly listed include total tumor volume, proportion enhancing, tumor location, ventricular invasion, multifocal or multicentric status, deep white matter invasion, and multiple MLS-related variables such as maximum MLS, edema crossing midline, enhancing tumor crossing midline, enlarged ventricles, and the level of maximum MLS (Rivera et al., 17 Feb 2026).

Text-quality evaluation is performed on HuskyBrain, a 184-case GBM cohort with pre-operative mpMRI, radiologist-authored reports, and tumor masks; a subset of 129 paired image–report cases is used for automated evaluation. The baselines are AutoRG-Brain and Seg-to-Exp. Metrics include BLEU-1, BLEU-2, ROUGE-1, ROUGE-2, BERTScore, RaTEScore, and TBFact score, precision, recall, and F1 (Rivera et al., 17 Feb 2026).

Metric	Best BTReport result	Baseline values reported
BLEU-1	0.248 ± 0.078, LLaMA3:70B V1	AutoRG-Brain 0.158 ± 0.072; Seg-to-Exp 0.085 ± 0.039
BLEU-2	0.136 ± 0.052, LLaMA3:70B V1	Lower than BTReport
ROUGE-1	0.371 ± 0.078, gpt-oss:120B V2	Lower than BTReport
ROUGE-2	0.115 ± 0.040, gpt-oss:120B V2	Lower than BTReport
TBFact Score	0.353 ± 0.151, gpt-oss:120B V2	0.072 for AutoRG; 0.014 for Seg-to-Exp
BERTScore	0.453 ± 0.055, gpt-oss:120B V2	0.327; 0.156
RaTEScore	0.577 ± 0.054, gpt-oss:120B V2	0.477; 0.409

Approximate randomization tests are reported to show BTReport variants are superior to baselines across all metrics with $p < 0.0001$. The qualitative examples in the appendix reinforce the same pattern. In a representative case with approximately 14 mm leftward midline shift and a large irregular enhancing right temporal lesion, BTReport correctly describes laterality, temporal or parietal involvement, mass effect, ventricular asymmetry, enhancement heterogeneity, necrotic core, and edema crossing midline, while slightly misestimating lesion dimensions relative to the ground-truth report. By contrast, AutoRG-Brain is reported to generate implausible dimensions, and Seg-to-Exp to produce text closer to an explanatory description than a clinically styled report (Rivera et al., 17 Feb 2026).

6. BTReport-BraTS, implementation stack, and auditability

The companion dataset, BTReport-BraTS, extends the framework from a report generator to a reusable resource for neuro-oncology RRG. It is built from the BraTS 2023 Adult Glioma dataset, combining training and validation splits for a total of $n = 1{,}470$ cases. For each case, the release includes pre-operative mpMRI, midline segmentations, all extracted metadata features, structured summary reports, and full BTReport-generated findings sections. The paper positions these reports as synthetic rather than clinician-authored and explicitly cautions that they should not be treated as clinical ground truth, but rather as high-quality pseudo-labels for open research use (Rivera et al., 17 Feb 2026).

The implementation stack is similarly modular. Preprocessing follows CaPTk. Registration and atlas propagation use SynthMorph. Pseudo-healthy anatomical segmentation uses SynthSeg. Tumor segmentations come from DeepMedic or the “Faking It” nnU-Net ensemble. VASARI-auto is modified to operate on subject-space segmentations and midline masks. Survival analysis uses the lifelines library. The evaluation stack includes TBFact with DeepSeek-R1, LangExtract for class–attribute extraction from reports, sentence-BERT for concept clustering, Gemma 3 27B for cluster summarization, and RadEval for BLEU, ROUGE, BERTScore, and RaTEScore. The paper also describes BTReview, a browser-based evaluation tool built around NiiVue for MRI visualization, segmentation overlays, MLS measurement, and structured radiologist survey collection (Rivera et al., 17 Feb 2026).

Interpretability is treated as a first-class property rather than a post hoc explanation layer. The framework claims direct traceability from each sentence to specific metadata entries and therefore to image-derived segmentations. Feature-level reliability is examined by comparing BTReport-extracted features against features extracted from ground-truth reports with LangExtract. Reported categorical accuracies include 1.00 for side of tumor epicenter, approximately 0.79–0.80 for cortical involvement, and approximately 0.52–0.56 for ventricular effacement. Reported numerical mean absolute errors include approximately 1.2–1.3 mm for midline shift, 0.41–1.48 lesions for number of lesions, and approximately 1.8–2.0 cm$^3$ for tumor volume. The paper interprets these results, together with TBFact and prompt-level constraints, as evidence that BTReport’s principal errors are more often attributable to segmentation or measurement discrepancies than to free-form hallucination (Rivera et al., 17 Feb 2026).

7. Limitations, generalizability, and position in the literature

BTReport’s limitations are explicit. The framework depends on segmentation quality, and ventricular effacement is identified as a particularly difficult feature, with accuracy around 0.5–0.56. The evaluation corpus, HuskyBrain, is a single-institution GBM cohort, and the survival analysis uses 461 BraTS’23 cases with complete metadata. The imaging scope is deliberately limited to structural MRI sequences T1, T1c, T2, and FLAIR; diffusion, perfusion, spectroscopy, and vascular imaging are excluded by design, which means findings tied to those modalities cannot be generated. In addition, BTReport currently generates only the Findings section, and its narrative style is tuned through in-context examples from a single institution (Rivera et al., 17 Feb 2026).

The authors nevertheless argue that the underlying methodology is site-agnostic and pathology-extensible provided that robust segmentation models and suitable feature ontologies are available. Proposed extensions include adding white matter hyperintensities, basal cistern status, stroke findings, and additional mpMRI modalities where available; using BTReview to collect radiologist feedback; and expanding BTReport-BraTS with additional labels and downstream task annotations. This suggests a roadmap in which deterministic feature interfaces remain fixed while segmentation models, ontologies, and prompting schemes evolve (Rivera et al., 17 Feb 2026).

Within the broader RRG literature, BTReport is positioned against chest-imaging systems trained on large paired image–report corpora and against neuro-oncology methods such as TextBraTS, AutoRG-Brain, and Seg-to-Exp. The paper’s stated distinction is not merely that it is two-stage, but that its intermediate representation is richer and clinically validated: VASARI descriptors, volumetrics, 3D midline shift, mass effect, and explicit anatomical relations are all made first-class conditioning variables for the LLM. BTReport is also presented as a neuro-oncology instantiation of a deterministic or hybrid philosophy associated with RadGPT, but adapted to the demands of tumor-distorted neuroanatomy through SynthMorph, SynthSeg, and specialized MLS estimation (Rivera et al., 17 Feb 2026).

In that sense, BTReport occupies an intermediate position between radiomics and report generation. Its features are chosen partly because they are commonly reported and partly because many are predictive of survival and molecular status. Its text output is not an opaque model extrapolation from pixels, but a formatted verbalization of an explicit structured representation. For researchers working on clinically grounded RRG in data-scarce domains, that architecture is the framework’s defining contribution (Rivera et al., 17 Feb 2026).