Desempeño en la detectabilidad de lesiones pulmonares metabólicas en sistema PET/CT aplicando técnicas de reconstrucciones gatilladas y deep learning

Abstract

Conventional imaging techniques such as computed tomography (CT) and positron emission tomography (PET) face significant challenges due to respiratory motion during PET/CT examinations, compromising accuracy in the identification and quantification of lung lesions. This study focuses on comparing the effectiveness of advanced image reconstruction techniques, including Ordered Subset Expectation Maximization (OSEM), Hyper Iterative Reconstruction (HYPER Iterative), and Deep Progressive Hyper Reconstruction (HYPER DPR) that integrates artificial intelligence, specifically in the detection of metabolic pulmonary micronodules. 35 oncology patients with malignant lesions in the lung were scanned using fluorodeoxyglucose labeled with Fluor-18 (18F-FDG). Images were acquired with varied reconstruction times and evaluated using contrast and signal-to-noise ratio (CNR) at different intervals. 55 micronodules were detected, and the techniques were evaluated for their ability to improve clarity and noise reduction. The results indicated that the HYPER DPR technique initially exhibited a higher CNR than OSEM and HYPER Iterative, although the latter showed significant improvements at longer intervals, outperforming HYPER DPR. The study concludes that the HYPER Iterative technique is the most effective for improving contrast-to-noise ratio over time, with HYPER DPR also demonstrating good performance at short intervals. These findings suggest that optimizing contrast against noise is possible through these advanced reconstruction techniques and highlight the effectiveness of acquisition times of 4 and 6 minutes as optimal for the quantification of lung lesions without the need for external respiratory devices.