Author: Jongduk Baek, Jooho Lee, Adam S. Wang 👨🔬
Affiliation: Stanford University, Yonsei University 🌍
Purpose: Optimizing the balance between radiation dose and image quality in computed tomography (CT) is important for minimizing patient X-ray exposure while maintaining diagnostic accuracy. While radiologist assessments are considered the gold standard for evaluating low-dose CT image quality, obtaining these scores is time-consuming and costly. This study investigates the use of negative log-likelihood (NLL), derived from deep generative models, as a proxy for radiologist scores in no-reference low-dose CT image quality assessment.
Methods: A score-based diffusion model was trained using normal-dose CT images from Mayo Clinic within a stochastic differential equation framework. Likelihood values were computed to evaluate how well an image aligns with the learned manifold of clean images, where deviations indicate lower image quality. Low-dose CT images were simulated at 25%, 50%, and 75% relative dose levels, with varying streak artifacts generated from sparse measurements of 64, 128, and 256 projection views. The LDCTIQAC 2023 dataset was used to assess the correlation between NLL and radiologists’ subjective scores. Additionally, Catphan phantom data was collected using a tabletop CBCT system with various X-ray tube currents to demonstrate practical utility.
Results: NLL effectively differentiated images based on noise and streak artifacts across varying dose levels. Strong correlations were observed between NLL and radiologist scores, with Pearson and Spearman correlation coefficients of 0.935 and 0.934, respectively. For the Catphan phantom data, NLL values showed consistency with traditional full-reference image quality metrics, such as PSNR and SSIM, providing a reliable measure.
Conclusion: Likelihood derived from deep generative models provides an effective surrogate for radiologist scores in low-dose CT image quality assessment. This offers a promising approach to emulate human perception without requiring human-labeled data.