🔎 Overview
LesionStudio is an open-source, GPU-accelerated Python platform that realistically inserts synthetic lesions of known ground truth directly into real clinical PET and SPECT acquisitions. It allows reconstruction algorithms to be evaluated under realistic patient-specific background heterogeneity, anatomy, and noise, capturing conditions that conventional phantoms cannot fully reproduce. Built on PyTomography, an open-source GPU-accelerated reconstruction engine developed in our Qurit lab, LesionStudio integrates lesion insertion, image reconstruction, and quantitative analysis within a single graphical workflow. The platform also includes a large collection of segmented real patient lesions that can be utilized.
In a validation study using FDG PET from diffuse large B-cell lymphoma (DLBCL) patients, a blinded physician review found inserted lesions visually indistinguishable from true disease (AUC=0.49), confirming the clinical realism of the framework's synthetic ground truth.
Clinical PET before and after lesion insertion. Red markers identify the inserted lesions.
🎯 Purpose
The platform enables task-specific optimization of reconstruction algorithms and parameters for both quantification and lesion-detection tasks. Lesions are placed directly in image space and mapped into raw acquisition data using a modality-specific forward projector, while stochastic background ablation removes the corresponding native signal so that ground truth remains fully known throughout reconstruction and analysis.
⚙️ Key Features
🧬 Real Lesion Library: A public library of 1,579 segmented FDG lymphoma lesions preserves authentic morphology and texture, complementing fully parametric spherical lesions for controlled recovery studies.
🖥️ Interactive GUI: The built-in viewer displays axial, sagittal, and coronal views simultaneously, with overlays that clearly indicate lesion placement. Once lesions are added, multiple reconstruction settings can be queued to run sequentially and automatically, eliminating the need for manual intervention between jobs.
📊 Automated quantitative analysis: Recovery, bias, noise-bias trade-off, contrast-to-noise ratio, edge artifacts, and the Detection Stability Index are all computed against known ground truth, with no separate analysis pipeline required.
🌐 Broadly applicable: Validated in FDG PET lymphoma imaging, the same framework extends naturally to theranostic SPECT dosimetry, long-axial-field-of-view PET, and objective evaluation of AI-based image enhancement.
🔓 Open-source and reproducible: Released under the MIT license with public source code, sample datasets, and the full lesion library, so results can be independently reproduced or extended to new diseases and radiopharmaceuticals.
💻 Supported Modalities & Algorithms
| Modality | Reconstruction Algorithms | Acquisition & Geometry |
PET | OSEM, BSREM (configurable iterations, subsets, regularization) | List-mode (.BLF) with correction files (.h5); single- and multi-bed |
SPECT | OSEM, BSREM, OSMAPOSL, KEM | DICOM projections; optional CT-based attenuation correction; lesion insertion via analytical forward projection or Monte Carlo (SIMIND) |
🔗 Access
💻 GitHub Repository: github.com/qurit/LesionStudio
📚 Citation
N. Aghakhanolia, J. Fowler, B. Panahi, A. Rahmim, P. E. Fernandez
A Realistic Lesion Insertion and Reconstruction Framework for Task-based Evaluations in Clinical SPECT and PET Imaging
Proc. SPIE Medical Imaging, Volume 13924: Physics of Medical Imaging; 139243P, 2026