InSilicoICH: Synthetic Intracranial Hemorrhage Modeling Tools

This repository contains tools for generating synthetic non-contrast CT datasets of intracranial hemorrhage (ICH). These datasets are designed to be used for developing, testing, and evaluating AI detection devices.

The Challenge: Data Scarcity in Medical AI

Intracranial hemorrhage (ICH) is a life-threatening brain bleed requiring immediate medical care. AI-powered Computer-Aided Triage (CADt) devices can accelerate diagnosis by analyzing emergency room CT scans (e.g., Rapid ICH K221456).

However, a significant data gap exists for pediatric patients. ICH occurs less frequently in children, leading to a scarcity of annotated pediatric data. An AI model trained predominantly on adults may perform poorly on pediatric scans, potentially delaying time-sensitive treatment for children.

Our Solution: In Silico Data Generation

To address this data availability challenge, InSilicoICH supplements real patient data with in silico (computer-simulated) data. By using realistic computational human phantoms and physics-based CT simulations, we can generate perfectly-labeled datasets at a fraction of the cost and time required for manual annotation of real data.

How It Works

Our simulation pipeline combines several state-of-the-art components:

1. Digital Phantoms: We use the MIDA phantom¹ and the NIHPD head phantoms ²³, which are detailed, anatomically-realistic patient-based models of the head.
2. Hemorrhage Insertion: A knowledge-based algorithm inserts synthetic hemorrhages into the phantoms. This algorithm controls the placement, shape, volume, and attenuation based on models from real hemorrhage segmentation data.
3. Physics-Based CT Simulation: The final phantom, complete with the synthetic hemorrhage, is imaged using XCIST, a realistic X-ray CT simulation framework that models the entire image acquisition process.

The tool currently supports three major hemorrhage subtypes:

• Intraparenchymal (IPH)
• Epidural (EDH)
• Subdural (SDH)

Intraventricular (IVH) and subarachnoid (SAH) hemorrhages are underdevelopment here.

Controllable Parameters

The simulation is highly customizable, allowing you to control dozens of parameters.

• Patient Characteristics

  • Phantom Identifier: Select a specific phantom from the cohort.
  • Age: 6.5 - 38 years (based on the available phantom library).

• Lesion Characteristics

  • Hemorrhage Type: IPH, SDH, EDH, or None.
  • Hemorrhage Volume: 0 - 100 mL.
  • Intensity (Attenuation): -30 - 100 HU.
  • Mass Effect: True or False.
  • Edema: 0 - 15 voxels (IPH only).

• Acquisition Characteristics

  • X-ray Tube Current: 10 - 1500 mA.
  • X-ray Tube Peak Voltage: 70 - 140 kVp.
  • Reconstruction FOV: 100 - 500 mm.
  • Reconstruction Kernel: e.g., Soft, Standard, Bone.
  • View Count: Number of projection views (e.g., 1000).

• Output & Metadata

  • Random Seed: For reproducibility.
  • Image & Mask Locations: Control where output files are saved.

Example Simulations

Below are example outputs using the MIDA phantom, showing the simulated CT images and the corresponding ground truth segmentation masks.

Simulated CT Images with ICH

Ground Truth Segmentation Masks

Tutorials and Usage

For worked examples and tutorials on how to use the included functions and tools to generate phantoms with ICH and all relevant parameters, please refer to the notebooks folder.

---

Installation and Setup

1. Prerequisites

• Python (tested on versions >=3.11 and <3.13)
• We recommend using a conda or venv virtual environment.

2. Install the Package

Install Directly

Install the package directly from GitHub using pip:

pip install git+https://github.com/DIDSR/InSilicoICH.git

Install from GitHub Clone

1. Clone the repository:

   git clone https://github.com/DIDSR/InSilicoICH.git
   cd InSilicoICH

2. Install dependencies:

   pip install .

References

Footnotes

1. U.S. Food and Drug Administration. (2023). MIDA: A Multimodal Imaging-Based Model of the Human Head and Neck (RST24NO05.01). https://cdrh-rst.fda.gov/mida-multimodal-imaging-based-model-human-head-and-neck ↩

2. VS Fonov, AC Evans, K Botteron, CR Almli, RC McKinstry, DL Collins and BDCG, Unbiased average age-appropriate atlases for pediatric studies, NeuroImage, In Press, ISSN 1053–8119, DOI: 10.1016/j.neuroimage.2010.07.033 ↩

3. VS Fonov, AC Evans, RC McKinstry, CR Almli and DL Collins Unbiased nonlinear average age-appropriate brain templates from birth to adulthood NeuroImage, Volume 47, Supplement 1, July 2009, Page S102 Organization for Human Brain Mapping 2009 Annual Meeting, DOI: 10.1016/S1053-8119(09)70884-5 ↩