A Pilot Study on Quantitative Accuracy and Radiomic Feature Stability of Deep Progressive Learning Reconstruction in...

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Radiology

Background:

• Deep progressive learning reconstruction (DPR) is a new deep learning algorithm for PET imaging.
• Its effects on quantitative metrics and radiomic feature stability are not fully understood.

Purpose of the Study:

• To systematically evaluate DPR against conventional ordered-subset expectation maximization (OSEM) in non-small cell lung cancer (NSCLC) PET imaging.
• To assess the impact of DPR on quantitative metrics (SUV, CNR, noise) and radiomic feature stability.

Main Methods:

• Retrospective analysis of 24 NSCLC patients' PET data.
• Reconstruction using OSEM and three DPR strength levels.
• Comparison of SUV, CNR, and background noise.
• Quantification of radiomic feature stability (93 features) using intra-patient coefficient of variation (COV_RF).

Main Results:

• DPR significantly increased SUV, especially in smaller tumors.
• Image quality decreased with higher DPR strengths; only the lowest DPR strength improved CNR.
• Radiomic feature stability varied: 33.3% were robust (median COV_RF ≤ 10%), while 40.9% were unstable.
• Commonly used features like glcm_Contrast were highly unstable.

Conclusions:

• DPR offers a trade-off between enhanced SUV quantification and image quality, necessitating parameter optimization.
• Radiomic feature stability is highly dependent on the reconstruction algorithm.
• Reliable use of advanced reconstruction techniques like DPR requires evidence-based selection of robust features.