Improving Lung Lesion Detection in Low Dose Positron Emission Tomography Images Using Machine Learning

Lung cancer suffers from poor prognosis, leading to high death rates. Combined PET/CT improves lung lesion detection but requires low dose protocols for frequent disease screening and monitoring. In this study, we investigate the feasibility of using machine learning to improve low dose PET images to standard dose, high-quality images for better lesion detection at low dose PET scans. We employ image quality transfer (IQT), which is a machine learning algorithm that uses patch-regression to map parameters from low to high-quality images e.g. enhancing resolution or information content. We acquired 20 standard dose PET images and simulated low dose PET images with 9 different count levels from the standard dose PET images. For each count levels, 10 pairs of standard dose PET images with one simulated low dose PET images were used to train linear, single non-linear regression tree, and random regression-forest models for IQT. The models were then used to estimate standard dose images from low dose images for each count levels for 10 different subjects. Improvement in image quality and lesion detection could be observed in the images estimated from the low dose images using IQT. Among the models employed, the regression tree model produced the best estimates of standard dose PET images. An average bias of less than 20% in SUV_mean of 25 lesions in the estimated images from the standard dose PET images can be obtained down to 7.5 × 10⁶ counts. Overall, despite the increase in bias, the improvement in image quality shows the potential of IQT in improving the accuracy in lesion detection.