kurtlab/BTReport-BraTS23

BTReport-BraTS23数据集是BTReport框架的配套数据集，旨在推进神经肿瘤学放射学报告生成（RRG）的研究。该数据集通过增强BraTS 2023影像数据集，添加了结构化的临床相关特征和合成放射学报告。BTReport提供了一种结构化的脑肿瘤报告方法，通过提取定量神经影像特征，并使用大型语言模型（LLMs）将其合成为专业的放射学报告。数据集包含标准化的VASARI特征（如增强、坏死、水肿）、3D中线移位定量估计（使用深度学习配准方法）、空间元数据（病变大小、坐标和解剖学涉及）以及由gpt-oss:120b和llama3:70b生成的放射学报告。数据集主要用于训练和微调LLMs用于医学报告生成、评估LLMs在确定性神经影像特征上的基础性，以及开发神经肿瘤学的自动临床文档工具。数据集不应用于初级临床诊断，且合成报告仅用于研究目的，不应替代人类认证的医疗记录。数据来源于BraTS 2023（脑肿瘤分割）挑战赛，不包含个人健康信息（PHI）。数据集存在模型偏见、技术限制和医学上下文限制等风险和局限性。

BTReport-BraTS23 is a companion dataset to the BTReport framework, designed to advance research in neuro-oncology radiology report generation (RRG). It augments the BraTS 2023 imaging dataset with structured, clinically relevant features and synthetic radiology reports. BTReport provides a structured approach to brain tumor reporting by extracting quantitative neuroimaging features and synthesizing them into professional radiology reports using Large Language Models (LLMs). The dataset includes standardized VASARI features (e.g., enhancement, necrosis, edema), 3D Midline Shift (quantitative estimation using deep learning registration), Spatial Metadata (lesion size, coordinates, and anatomical involvement), and radiology reports generated by gpt-oss:120b and llama3:70b. The dataset is intended for training and fine-tuning LLMs for medical report generation, evaluating the grounding of LLMs on deterministic neuroimaging features, and developing automated clinical documentation tools for neuro-oncology. It should not be used for primary clinical diagnosis, and synthetic reports are for research purposes only, not to replace human-certified medical records. The data is derived from the BraTS 2023 (Brain Tumor Segmentation) challenge and contains no PHI. The dataset has risks and limitations including model bias, technical limitations, and medical context constraints.