Preserving Diagnostic Integrity in Ultra-Low-Dose Lung CT via Domain-Specific Perceptual Guidance
Imaging Research, Sheba Medical Center, Israel
Background: Ultra-low-dose computed tomography (ULDCT) at ~5% of standard radiation dose minimizes cumulative radiation exposure for patients. However, severe quantum noise and streak artifacts compromise image quality and spatial resolution, making diagnostic evaluation challenging. Conventional deep-learning denoising algorithms often over-smooth subtle pulmonary textures or introduce misleading artificial artifacts. We evaluated a clinical-semantic denoising framework that utilizes a specialized perceptual guidance model pre-trained on computed tomography (CT) data to preserve diagnostic utility.
Methods: Paired ULDCT and standard-dose CT lung scans were spatially aligned using deformable image registration. Feature-space representations from a perceptual model—pre-trained on large-scale clinical CT data—were integrated into the training loss functions of various generative network architectures. The reconstructed images were evaluated on direct visual quality metrics and on automated clinical tools, including lung segmentation, airway (bronchi) segmentation, emphysema segmentation, and emphysema percentage quantification. Qualitative evaluation by expert radiologists is currently a work in progress (WIP).
Results: Images processed with the CT-trained perceptual guidance demonstrated marked performance gains across all automated clinical tasks compared to standard denoising baselines. The proposed approach achieved significantly higher structural overlap and anatomical accuracy for lung, airway, and emphysematous tissue segmentations. Furthermore, it yielded a stronger correlation in emphysema percentage quantification relative to standard-dose reference scans. Preliminary visual inspection indicates superior preservation of fine pulmonary parenchyma without artificial smoothing, with formal radiological scoring ongoing.
Conclusions: Incorporating domain-specific perceptual guidance pre-trained on CT data significantly enhances the performance of automated diagnostic tools on 5%-dose CT scans. Finalizing radiological validation will further confirm its clinical utility for both human interpretation and automated analysis.