Deep virtual CT workflow for evaluating AI in low-dose CT lung cancer screening: a 2D demonstration

Over the past decade, most CT systems in the United States have been cleared by the FDA with an indication for lung cancer screening (LCS) using specialized low-dose CT (LDCT) acquisition protocols. At the time of these clearances, CT systems were primarily equipped with traditional filtered backprojection (FBP) and statistical/model-based iterative reconstruction (IR) methods. Nowadays, deep learning (DL)-based reconstruction and denoising (DLR/D) methods are widely available for LDCT, with the potential to generate visually appealing images at reduced dose levels. Yet, the issues of generalizability and instability are recognized, and the stability and fidelity (non-hallucination) must be assessed for ultra-low-dose CT scans to ensure that they preserve diagnostic image quality for various patient characteristics and pathological conditions, and are safe for use. These assessments are likely to require a huge amount of testing data, which may not be possible with patient scans. In this work, we study the feasibility of using a deep learning-based virtual CT workflow to evaluate DLR/D methods in LDCT for LCS. To achieve this goal, our deep virtual CT workflow is designed with the four major components: AI-generated patients, a realistic simulated CT sinogram, deep reconstruction, and performance evaluation in terms of both traditional image quality metrics and AI-based lung nodule detection accuracy. Preliminary results are presented to demonstrate the potential of such a deep workflow for evaluating DLR/D methods in LDCT LCS.