Validation of low-dose lung cancer PET-CT protocol and PET image improvement using machine learning

Area of Science:

• Nuclear Medicine
• Medical Imaging
• Machine Learning Applications in Radiology

Background:

• Lung cancer screening requires frequent imaging, necessitating low-dose protocols to minimize patient radiation exposure.
• Low-dose PET-CT scans often suffer from reduced image quality and statistics, potentially impacting lesion detection.
• Machine learning (ML) offers potential for enhancing low-statistics medical images.

Purpose of the Study:

• To evaluate a simplified low-dose (LD) PET-CT protocol for lung screening, using only 30% of the standard effective dose.
• To investigate the feasibility of improving the clinical utility of low-statistics scans via machine learning techniques.
• To assess lesion detection performance using the developed LD PET-CT protocol.

Main Methods:

• Acquired standard-dose (SD) PET images and derived actual LD (ALD) and simulated LD (SLD) PET images at various count levels.
• Employed image quality transfer (IQT), an ML algorithm, to map low-quality to high-quality image parameters.
• Trained and applied IQT models (global linear, single tree, random forest) using patch-regression on SD/SLD PET image pairs to estimate SD images from LD images.

Main Results:

• The LD PET-CT protocol achieved high lesion detectability with a sensitivity of 0.98 and specificity of 1.
• A random forest ML model enabled an additional 11.7% dose reduction without compromising lesion detectability.
• Significant underestimation of Standardized Uptake Value (SUV) by 30% was observed with the further dose reduction.

Conclusions:

• The LD PET-CT protocol is validated for effective lesion detection in lung screening using ALD PET scans.
• Machine learning methods can substantially improve image quality or allow further dose reduction while preserving clinical value.
• SUV quantification bias necessitates protocol adjustments for clinical implementation of ML-enhanced LD PET-CT.