Exploring the Impact of Acquisition and Reconstruction Parameters on an Imaging-Based Lung Cancer Risk Model

Image-based risk models have the potential to aid in the identification of individuals who would benefit from cancer screening. However, models need to be robust against variations in image acquisition and reconstruction parameters, which alter the appearance of images and may lead to different downstream predictions. We evaluated Sybil, an imaging-based lung cancer model that predicts up to six-year risk, on a lung cancer screening dataset that was acquired and reconstructed using a range of parameters. Using raw projection data from 169 retrospectively acquired low-dose computed tomography (LDCT) scans, we generated six image conditions for each case, varying reconstruction kernels (smooth, medium, sharp) and slice thicknesses (1.0mm, 2.0mm). Each image condition was processed and run through the pre-trained Sybil model in the same way. Variations in predicted risk scores were observed across various kernels and slice thicknesses, suggesting that deep features derived from LDCT scans can be sensitive to nuanced variations in acquisition and reconstruction parameters. Our study underscores the importance of enhancing understanding of how technical parameters impact predictive models like Sybil to enhance the reliability of model outputs, facilitating more accurate and robust clinical decision support.