Deep learning image enhancement algorithms in PET/CT imaging: a phantom and sarcoma patient radiomic evaluation

Area of Science:

• Medical Imaging
• Radiology
• Artificial Intelligence in Medicine

Background:

• Positron Emission Tomography/Computed Tomography (PET/CT) imaging offers quantitative data for tumor characterization.
• Deep learning (DL) techniques are increasingly used for denoising PET data, requiring robust clinical evaluation.
• Quantitative image assessment using radiomic features can further evaluate DL algorithms.

Purpose of the Study:

• To compare two manufacturer deep-learning (DL) image enhancement algorithms against gold-standard reconstruction techniques.
• To assess the performance of DL algorithms using radiomic features in phantom and sarcoma patient data.
• To evaluate the potential of DL in harmonizing radiomics and for quantitative assessment of DL algorithms.

Main Methods:

• Utilized retrospective [18F]FDG PET/CT sarcoma patient data from GE Discovery 690/710 scanners.
• Employed a modular heterogeneous imaging phantom with repeat acquisitions for phantom studies.
• Compared DL-enhanced images to gold-standard reconstructions and algorithm input images using 93 International Biomarker Standardization Initiative (IBSI) radiomic features.

Main Results:

• A small percentage of radiomic features (4.0% phantom, 9.7% patient) differed significantly between DL-enhanced and gold-standard images.
• Larger differences were observed between DL-enhanced and algorithm input images (29.8% phantom, 43.0% patient).
• Over 80% of radiomic features showed no significant difference when comparing DL-enhanced to gold-standard images.

Conclusions:

• DL-enhanced PET/CT images closely resemble gold-standard reconstructions, as evidenced by radiomic feature analysis.
• DL algorithms demonstrate potential for harmonizing radiomic data across different imaging protocols.
• Radiomic features are effective for the quantitative evaluation of DL-based image enhancement in PET/CT.