Fully 3D Active Surface with Machine Learning for PET Image Segmentation.

Area of Science:

• Medical Imaging
• Computational Biology
• Machine Learning

Background:

• Current 3D tumor volume reconstruction from Positron Emission Tomography (PET) images often relies on segmenting individual 2D slices, neglecting valuable cross-slice information.
• Existing 2D segmentation methods can be suboptimal for accurate volumetric analysis.

Purpose of the Study:

• To develop and evaluate a novel 3D active surface algorithm for direct three-dimensional tumor volume reconstruction from PET images.
• To enhance segmentation accuracy by integrating a machine learning component into the 3D active surface model.

Main Methods:

• A 3D active surface algorithm was developed to segment entire PET image stacks simultaneously, handling topological changes naturally.
• A machine learning component, based on discriminant analysis, was incorporated as a forcing term within the active surface's energy function.
• The algorithm was trained and tested on PET data from patients with lung, head and neck, and brain cancers.

Main Results:

• The 3D active surface algorithm demonstrated superior performance compared to traditional 2D active contour methods across all investigated anatomical regions.
• Achieved high Dice Similarity Coefficients: 90.47 ± 2.36% for lung, 88.30 ± 2.89% for head and neck, and 90.29 ± 2.52% for brain cancer.
• The machine learning component required minimal training data, offering an efficient enhancement to segmentation accuracy.

Conclusions:

• The developed 3D active surface algorithm provides a practical and significant benefit for PET image segmentation and tumor volume reconstruction.
• Migrating from 2D to a 3D segmentation system offers substantial improvements in accuracy and efficiency for oncological imaging analysis.