A Framework for Synthetic ICH to Assess AI Generalizability

Synthetic intracranial hemorrhage (ICH) datasets and artificial intelligence model files to recreate figures found in Knowledge-based synthetic intracranial hemorrhage CT datasets for device evaluations and generalizability assessments (under submission) using the Jupyter Notebooks found in the manuscript repository.

This repository contains the dataset used to reproduce the results presented in the paper: Knowledge-based synthetic intracranial hemorrhage CT datasets for device evaluations and generalizability assessments (preprint in Files).

Deep learning models for Computer-Assisted Detection (CAD) of intracranial hemorrhage (ICH) often struggle with generalizability when encountering CT data with characteristics underrepresented in their training sets (e.g., variations in patient demographics, hemorrhage types, or image acquisition parameters).

This project introduces an open-source framework to:

Generate synthetic ICH CT data by inserting realistic, modeled hemorrhages (epidural, subdural, intraparenchymal) into a digital head phantom.

Simulate mass effect and control hemorrhage volume and attenuation based on real data distributions.

Create datasets with varied CT acquisition parameters (mAs, kVp) to robustly evaluate the generalizability of ICH detection models.

Our work validates this approach by demonstrating comparable performance of an ICH detection model on our synthetic dataset (AUC 0.877) versus an independent real dataset (AUC 0.919). This framework enables more comprehensive testing and evaluation of CAD devices for ICH.

本合成颅内出血数据集与人工智能模型文件，可借助稿件仓库中的Jupyter Notebooks，复现已投稿论文《Knowledge-based synthetic intracranial hemorrhage CT datasets for device evaluations and generalizability assessments》中的图表。

本仓库包含用于复现论文《Knowledge-based synthetic intracranial hemorrhage CT datasets for device evaluations and generalizability assessments》（预印本存于Files目录）中实验结果的数据集。

用于颅内出血（intracranial hemorrhage, ICH）计算机辅助检测（Computer-Assisted Detection, CAD）的深度学习模型，在遇到训练集中特征表征不足的CT数据时，往往难以保证泛化能力——例如患者人口统计学特征、出血类型或图像采集参数的差异。

本项目提出了一款开源框架，可实现以下功能：
1. 通过将符合真实建模的出血（硬膜外出血、硬膜下出血、脑实质内出血）植入数字化头部体模，生成合成颅内出血CT数据；
2. 基于真实数据分布模拟占位效应，并调控出血量与CT衰减值；
3. 构建包含不同CT采集参数（管电流时间积mAs、管电压kVp）的数据集，以稳健评估颅内出血检测模型的泛化能力。

本研究通过验证发现，颅内出血检测模型在本项目合成数据集上的曲线下面积（Area Under Curve, AUC）为0.877，与独立真实数据集上的AUC 0.919性能相当，从而证实了该框架的有效性。该框架可用于更全面地测试与评估用于颅内出血检测的计算机辅助检测设备。