BTReport: A Framework for Brain Tumor Radiology Report Generation with Clinically Relevant Features

TL;DR

BTReport introduces a two-stage framework for brain tumor radiology report generation that deterministically extracts clinically relevant imaging features and then leverages large language models to generate structured, interpretable reports. By grounding interpretation in features such as VASARI descriptors and 3D midline shift, the approach improves factual accuracy and reduces hallucinations without task-specific model fine-tuning. The authors validate feature relevance through clustering of real radiology concepts and survival analyses, and demonstrate superior lexical and factual alignment against baselines on HuskyBrain BraTS-derived data. To facilitate research, BTReport-BraTS provides an open companion dataset and BTReview enables radiologist-based evaluation of generated reports, advancing practical, clinically aligned RRG in neuro-oncology.

Abstract

Recent advances in radiology report generation (RRG) have been driven by large paired image-text datasets; however, progress in neuro-oncology has been limited due to a lack of open paired image-report datasets. Here, we introduce BTReport, an open-source framework for brain tumor RRG that constructs natural language radiology reports using deterministically extracted imaging features. Unlike existing approaches that rely on large general-purpose or fine-tuned vision-language models for both image interpretation and report composition, BTReport performs deterministic feature extraction for image analysis and uses large language models only for syntactic structuring and narrative formatting. By separating RRG into a deterministic feature extraction step and a report generation step, the generated reports are completely interpretable and less prone to hallucinations. We show that the features used for report generation are predictive of key clinical outcomes, including survival and IDH mutation status, and reports generated by BTReport are more closely aligned with reference clinical reports than existing baselines for RRG. Finally, we introduce BTReport-BraTS, a companion dataset that augments BraTS imaging with synthetically generated radiology reports produced with BTReport. Code for this project can be found at https://github.com/KurtLabUW/BTReport.

Figures (6)

• Figure 1: BTReport Overview: Interpretable, clinically meaningful variables are deterministically extracted for each case, including demographics, VASARI features, and 3D midline shift measurements. These features are utilized by context-guided LLMs for clinically grounded radiology report generation.
• Figure 2: Atlas-based 3D midline shift (MLS) estimation using SynthMorph synthmorph, in which atlas midline annotations are registered to patient imaging and voxel-wise distances to an ideal midline axis are computed per axial slice.
• Figure 3: Kaplan-Meier (KM) plots for demographic and radiogenomic features, and deterministic features extracted with the BTReport framework. Survival probabilities at specific time points are obtained by projecting vertically from the time of interest to the curve and horizontally to the y-axis. Example: for the Age feature, at 500 days post-diagnosis, the estimated survival probability for the older age group (High) is $\sim$ 20%, compared to $\sim$ 50% for the younger age group (Low). Kaplan–Meier analyses indicate that many of these features are predictive of overall survival, highlighting their clinical relevance and motivating their use as structured inputs for radiology report generation.
• Figure 4: Visualization of features extracted from free text radiologist-authored report using LangExtract.
• Figure 5: Visualization of features extracted from free text BTReport-generated report using LangExtract.