AIRPET: Virtual Positron Emission Tomography

This paper introduces AIRPET, a web-based, AI-assisted platform that streamlines the complex, multi-stage process of PET scanner design and evaluation by integrating detector simulation, image reconstruction, and LLM-driven geometry creation into a single accessible workflow.

The Gist

Imagine you want to build a brand-new, high-tech camera that can see inside the human body to spot diseases. This isn't a regular camera; it's a PET scanner, a device that detects tiny particles to create 3D pictures of how your body's organs are working.

Building one of these machines is incredibly hard. It's like trying to build a Ferrari, drive it on a race track, and then act as the doctor diagnosing the driver's health—all by yourself. The paper introduces AIRPET, a new web-based tool designed to make this impossible job much easier by acting as a "co-pilot" for scientists.

Here is how AIRPET works, broken down into simple steps:

1. The Problem: A Three-Headed Monster

Currently, designing a PET scanner is split into three very difficult jobs that usually require different experts:

• The Architect: Designs the physical shape and materials of the detector (using complex math and physics software).
• The Simulator: Runs a "virtual test drive" to see how the machine would behave if it existed.
• The Doctor: Looks at the resulting images to see if they are clear enough to diagnose a patient.

Most researchers can only do one of these jobs well. If you are great at design, you might struggle with the simulation code. If you are a coder, you might not understand the medical side. This creates a "silo" where experts can't easily talk to each other.

2. The Solution: AIRPET (The All-in-One Workshop)

AIRPET is a website that brings all three of these jobs into a single, easy-to-use workshop. It's like a "Lego set" for PET scanners, but with a smart robot helper.

• The Smart Robot (AI Assistant): Instead of writing hundreds of lines of confusing code to describe your machine, you can just type a request to the AI. You might say, "Build a ring of 16 crystals with a 90cm radius." The AI acts as a translator, turning your simple words into the complex technical files the computer needs to build the virtual machine.
• The Virtual Test Drive (Simulation): Once the machine is "built" in the computer, AIRPET runs a simulation. It shoots virtual particles through your design to see how they bounce around, just like crash-testing a car in a video game before building the real thing.
• The Picture Maker (Reconstruction): After the simulation, the tool takes the data and instantly turns it into a 3D picture. You can see if your design actually produces a clear image or if it's blurry.

3. A Real Example: The "CRYSP" Test

The authors tested this tool using a specific design called CRYSP. They used AIRPET to build a virtual scanner made of special crystals. They placed a "phantom" (a fake body part made of water with six tiny balls inside) in the center.

They told the computer to simulate the scanner looking at these balls. Within minutes, AIRPET generated a 3D image showing the six balls clearly. This proved that the tool can take a design idea, simulate it, and show you the result without needing a team of ten experts.

4. What's Next? (The Future Workshop)

The paper explains that AIRPET is still under construction, like a house that has the walls up but needs more furniture. The authors plan to add:

• Better AI Tools: Instead of just asking the AI to "write code," they want to give the AI specific "tools" (like a pre-made function to arrange crystals in a circle) so it makes fewer mistakes.
• A Library of Parts: A digital shelf where users can grab pre-made parts (like standard medical test objects) instead of building everything from scratch.
• The "AI Doctor": Eventually, they want to add an AI that can look at the generated 3D images and give a second opinion on how good the image is, acting as a training partner for real doctors.

The Bottom Line

AIRPET is a web-based platform that uses Artificial Intelligence to help scientists design, test, and visualize PET scanners all in one place. It lowers the barrier to entry, allowing smaller teams or individuals to experiment with new scanner designs without needing to be masters of every single step of the process. It is currently a research tool for building better machines, not a device used directly on patients yet.

Technical Summary

Technical Summary: AIRPET – Virtual Positron Emission Tomography

Problem Statement
The design and evaluation of new Positron Emission Tomography (PET) scanner technologies is a complex, multidisciplinary process typically divided into three distinct stages: (1) detector design and simulation, (2) image reconstruction, and (3) image interpretation. Each stage demands specialized expertise, creating a significant barrier to entry for researchers and engineers who cannot easily manage all three simultaneously. Furthermore, the current landscape relies on a variety of specialized software tools that operate in silos, requiring significant effort to integrate and often preventing experts in one domain from fully appreciating the implications of their work on other stages. While medical professionals perform the final interpretation, there is a need for tools that can assist in generating training data and providing an "AI partner" to compare diagnoses against radiologists.

Methodology
To address these challenges, the authors developed AIRPET (AI-driven Revolution in Positron Emission Tomography), an open-source, web-based platform that integrates the entire PET design workflow into a single interface. The system operates on a client-server model:

• Frontend: Built with HTML/JavaScript and three.js, it provides a visual editor for 3D geometry manipulation.
• Backend: Built with Python/Flask, it manages project state, simulation execution, and image reconstruction.

The workflow proceeds through the following technical components:

1. Detector Design: Users define geometry via a visual editor or text-based prompts. The system utilizes an internal JSON representation and supports the standard Geant4 GDML format for import/export. A key feature is the integration of Large Language Models (LLMs), currently supporting locally hosted open-source models (via Ollama) and cloud-based models (Google Gemini). Users can input natural language commands (e.g., "create a minimal PET geometry...") which the AI attempts to translate into the necessary JSON structure. The system also supports CAD import via STEP files using the pythonocc library.
2. Detector Simulation: Upon user initiation, the backend generates GDML and Geant4 macro (.mac) files based on the defined geometry and parameters. The simulation runs in a separate thread, providing real-time progress updates. Output data, including detector hits and particle tracks, is stored in HDF5 format.
3. Image Reconstruction: The system processes simulated 511 keV gamma ray interactions into Lines of Response (LORs). It applies basic detector physics effects, such as Gaussian smearing for position and energy resolution. The platform interfaces with the parallelproj library to perform 3D image reconstruction using the Maximum Likelihood Expectation Maximization (MLEM) algorithm.
4. Medical Evaluation (Proposed): The authors outline a future module employing multi-modal LLMs to provide qualitative assessments of reconstructed images, intended as a training tool rather than a substitute for professional medical diagnosis.

Key Contributions and Results

• Unified Workflow: AIRPET successfully bridges the gap between detector design, simulation, and image reconstruction, allowing users to move from geometry definition to a visualized 3D PET image within a single web interface.
• AI-Assisted Geometry Creation: The platform demonstrates the feasibility of using LLMs to generate complex detector geometries from text prompts, significantly reducing the manual coding burden, though the authors note this currently requires human fine-tuning.
• Demonstration Case (CRYSP): The paper presents a validation case modeling a simplified version of the CRYSP1M total-body PET scanner. The system simulated a 6-sphere water-filled phantom with varying radii and activity ratios. Using 108 positron-electron decays and 10 iterations of the MLEM algorithm, AIRPET successfully reconstructed a 3D image of the phantom, visualizing the activity distribution. The normalization was performed using a sensitivity matrix derived from simulated random LORs.

Significance and Claims
The authors position AIRPET as a tool to democratize PET research and development by lowering the technical barriers associated with integrating simulation and reconstruction tools. The platform aims to serve two primary functions:

1. A Training Ground: By generating training sets and simulated realistic problems, it aids in the education of researchers and the development of AI diagnostic tools.
2. An AI Partner: It provides a mechanism to compare AI-generated diagnoses with those of radiologists, potentially accelerating the evaluation of new scanner designs.

The paper maintains a modest tone regarding its current capabilities, acknowledging that the AI-assisted geometry creation is not yet a replacement for skilled human designers and that the medical evaluation module is a future goal rather than a current feature. The primary value proposition is the integration of these disparate stages into a cohesive, accessible workflow to facilitate more efficient PET scanner design.