Related Experiment Video
Updated: Jun 18, 2026
Machine Learning Algorithms for Early Detection of Bone Metastases in an Experimental Rat Model
Published on: August 16, 2020
An iterative technique to segment PET lesions using a Monte Carlo based mathematical model
S A Nehmeh1, H El-Zeftawy, C Greco
1Department of Medical Physics, Memorial Sloan-Kettering Cancer Center, New York, New York 10021, USA. nehmehs@mskcc.org
Purpose:
The need for an accurate lesion segmentation tool in 18FDG PET is a prerequisite for the estimation of lesion response to therapy, for radionuclide dosimetry, and for the application of 18FDG PET to radiotherapy planning. In this work, the authors have developed an iterative method based on a mathematical fit deduced from Monte Carlo simulations to estimate tumor segmentation thresholds.
Methods:
The GATE software, a GEANT4 based Monte Carlo tool, was used to model the GE Advance PET scanner geometry. Spheres ranging between 1 and 6 cm in diameters were simulated in a 10 cm high and 11 cm in diameter cylinder. The spheres were filled with water-equivalent density and simulated in both water and lung equivalent background. The simulations were performed with an infinite, 8/1, and 4/1 target-to-background ratio (T/B). A mathematical fit describing the correlation between the lesion volume and the corresponding optimum threshold value was then deduced through analysis of the reconstructed images. An iterative method, based on this mathematical fit, was developed to determine the optimum threshold value. The effects of the lesion volume and T/B on the threshold value were investigated. This method was evaluated experimentally using the NEMA NU2-2001 IEC phantom, the ACNP cardiac phantom, a randomly deformed aluminum can, and a spheroidal shape phantom implemented artificially in the lung, liver, and brain of patient PET images. Clinically, the algorithm was evaluated in six lesions from five patients. Clinical results were compared to CT volumes.
Results:
This mathematical fit predicts an existing relationship between the PET lesion size and the percent of maximum activity concentration within the target volume (or threshold). It also showed a dependence of the threshold value on the T/B, which could be eliminated by background subtraction. In the phantom studies, the volumes of the segmented PET targets in the PET images were within 10% of the nominal ones. Clinically, the PET target volumes were also within 10% of those measured from CT images.
Conclusions:
This iterative algorithm enabled accurately segment PET lesions, independently of their contrast value.
