AIRPET: Virtual Positron Emission Tomography

Positron Emission Tomography (PET) is a powerful medical imaging technique, but the design and evaluation of new PET scanner technologies present significant challenges. The process is typically divided into three major stages: 1. detector design and simulation, 2. image reconstruction, and 3. image interpretation. Each of these stages requires significant expertise, making it difficult for individuals or small teams to manage all three at once. AIRPET (AI-driven Revolution in Positron Emission Tomography) is a web-based platform designed to address this challenge by integrating all phases of PET design into a single, accessible, and AI-assisted workflow. AIRPET provides an interface to large language models (LLMs) for assisted geometry creation and an interface for basic PET image reconstruction with the potential for further expansion. Here we introduce AIRPET and outline its current functionality and proposed additions.

💡 Research Summary

The paper presents AIRPET (AI‑driven Revolution in Positron Emission Tomography), an open‑source, web‑based platform that unifies the three traditional stages of PET scanner development—detector design and simulation, image reconstruction, and medical interpretation—into a single, AI‑assisted workflow. The authors argue that each stage normally requires specialized software and expertise, creating a high barrier for small research groups. AIRPET addresses this by providing a graphical front‑end for geometry creation, automatic integration with the Geant4 Monte‑Carlo toolkit for physics simulation, and a reconstruction back‑end based on the parallelproj library.

Platform architecture: The system follows a client‑server model. The front‑end, built with HTML, JavaScript and three.js, offers an interactive 3D editor where users can add, delete, or modify solids, materials, logical and physical volumes, and export/import GDML files. Geometry data are stored internally as JSON for version control. The back‑end runs on Python/Flask, handling project state, generating Geant4 macro files, launching simulations in separate threads, and storing results in HDF5 format.

AI‑assisted geometry creation: AIRPET integrates large language models (LLMs) via locally hosted Ollama models or cloud‑based Google Gemini. Users can type natural‑language commands such as “create a minimal PET geometry consisting of a ring of 16 crystals with a 90 mm radius.” The LLM parses the request and produces the corresponding JSON structure, which is then translated into GDML. While the current implementation does not replace a skilled designer, it dramatically speeds up the initial drafting phase, after which users can manually refine the output. The system also supports CAD import through pythonocc, converting STEP files into tessellated Geant4 solids.

Simulation workflow: Upon user request, the back‑end writes the current geometry to a GDML file, creates a Geant4 macro (.mac) reflecting the defined source distribution and simulation parameters, and launches Geant4. Progress logs are streamed back to the web UI. The simulation records detector hits and particle tracks, saving them in HDF5 for downstream analysis. At present, detector physics effects are limited to Gaussian smearing of hit positions and energies, but the architecture allows future inclusion of time‑of‑flight (TOF), scatter, and energy‑window modeling.

Image reconstruction: The platform extracts lines of response (LORs) from simulated hits, applying the same Gaussian smearing to emulate realistic position resolution. Reconstruction is performed with the parallelproj library using the Maximum Likelihood Expectation Maximization (MLEM) algorithm. Users can set the number of iterations; the resulting 3‑D image volume is displayed slice‑by‑slice directly in the browser. Planned extensions include filtered back‑projection (FBP), TOF‑aware reconstruction, scatter correction, and additional regularization techniques.

Medical evaluation (future work): A forthcoming module will employ multimodal LLMs (e.g., MedGemma, LLaVA‑Med) to provide qualitative assessments of reconstructed PET images. The goal is not to replace radiologists but to give designers rapid feedback on how geometry choices affect diagnostic image quality.

Demonstrative use case: The authors model a simplified version of the CRYSP total‑body PET scanner, consisting of CsI crystals arranged in rings of ~40 cm radius. A six‑sphere water phantom with activity ratios of 4:1 (sphere:water) is placed at the center. Ten million positron decays are simulated, and a 10‑iteration MLEM reconstruction yields an image that correctly resolves the spheres. Normalization uses a sensitivity matrix derived from 20 million random LORs and applies water attenuation (μ = 0.096 cm⁻¹).

Future roadmap: The paper outlines several enhancements: adding alternative reconstruction algorithms (FBP, TOF, scatter corrections), expanding AI “tool use” by exposing parameterized functions for geometry generation, building libraries of standard PET components and medical phantoms (e.g., Jaszczak phantom), and integrating pre‑trained medical image evaluation models for automated quality metrics.

Open‑source availability: AIRPET is released under the MIT license on GitHub (https://github.com/jerenner/airpet), encouraging community contributions, plugin development, and broader adoption.

In summary, AIRPET offers a cohesive, web‑accessible environment that lowers the entry barrier for PET scanner research. By coupling intuitive geometry editing, automated Geant4 simulation, and AI‑enhanced workflow assistance, it enables rapid prototyping, immediate visual feedback, and a closed loop between design decisions and imaging performance. The proposed extensions promise to further enrich the platform, making it a potentially central hub for future PET development and education.