Comparison of Automatic Segmentation and Preprocessing Approaches for Dynamic Total-Body 3D Pet Images with Different...

Area of Science:

• Medical imaging
• Image analysis
• Computational science

Background:

• Automatic segmentation is crucial for PET image analysis, but few tools exist for modern large datasets.
• Existing automatic methods are often outdated, designed for smaller images, and lack available implementations.

Purpose of the Study:

• To evaluate the feasibility of commonly used unsupervised segmentation methods on large, modern total-body PET images.
• To identify promising algorithms for future development in automated PET image segmentation.

Main Methods:

• Tested 17 unsupervised segmentation algorithms on dynamic total-body PET images from five datasets.
• Included various preprocessing techniques and segmentation methods.
• Validated results against manual segmentations using Jaccard index, Dice score, precision, and recall.

Main Results:

• Only 6 out of 17 methods were computationally feasible for large PET images.
• Hierarchical clustering and HDBSCAN showed the lowest performance.
• Gaussian Mixture Models (GMM) and k-means clustering achieved median Jaccard indices of 0.58, with GMM outperforming k-means on human data.

Conclusions:

• Most unsupervised segmentation methods are computationally infeasible for modern PET imaging.
• Gaussian Mixture Models and k-means are the most promising candidates for further development.
• Preprocessing, especially denoising, improved results, but segmenting small organs remains challenging.