Background based Gaussian mixture model lesion segmentation in PET.

Area of Science:

• Medical Imaging
• Nuclear Medicine
• Computational Biology

Background:

• Quantitative (18)F-fluorodeoxyglucose positron emission tomography (FDG-PET) imaging faces challenges in accurate lesion delineation due to low signal-to-noise ratio (SNR), resolution limits, and partial volume effects.
• These limitations significantly impact cancer assessment, treatment planning, and patient follow-up, necessitating improved segmentation techniques.

Purpose of the Study:

• To develop and validate a novel statistical clustering segmentation algorithm for enhanced lesion delineation in FDG-PET imaging.
• To improve the robustness of segmentation against variations in lesion size, noise, and contrast by incorporating background features and contiguity priors.

Main Methods:

• A modified eight-class Gaussian mixture model (GMM) clustering algorithm was employed, constraining background class parameters via prior analysis (background modeling).
• Expectation maximization was focused on lesion detection classes, with an additional variant incorporating spatial priors to enhance segmentation of connected objects.
• The algorithm was tested on simulated, phantom, and clinical datasets, with comparisons against standard GMM and state-of-the-art methods using metrics like Volume Error (VE) and Dice index.

Main Results:

• The proposed GMM segmentation with background modeling demonstrated superior performance over standard GMM and other methods across simulated, phantom, and clinical data.
• Achieved median accuracy indexes included VE <3%, Dice >0.88 in simulations, and VE <23%, Dice >0.74 in phantom data.
• The algorithm showed robustness to image statistic variations and initialization, with spatial priors further improving accuracy in heterogeneous backgrounds (median VE decreased from 13% to 7%).

Conclusions:

• Constraining statistical clustering algorithms, particularly with background modeling, significantly enhances their performance in FDG-PET lesion segmentation.
• The inclusion of spatial priors is beneficial for segmenting lesions in complex, heterogeneous backgrounds.
• The developed algorithm's robustness and accuracy support its clinical applicability for improved oncological imaging analysis.