Improving visibility of small anatomical details on low and ultra-low dose computed tomography with artificial intelligence-based image reconstructions

To assess how computed tomography (CT) image reconstruction techniques affect perceived diagnostic image quality at varying radiation dose levels in chest imaging. A PBU-50 anthropomorphic phantom (small adult-sized model) and an air-dried human lung specimen were scanned on the same CT system (Revolution Apex™, GE Healthcare) at six dose levels (CTDIvol) from 0.07 to 2.19 mGy for the smallest phantom size. Images were reconstructed using deep learning image reconstruction-high (DLIR-H), adaptive statistical iterative reconstruction at 40 per cent (ASiR-V), and filtered back projection (FBP). Five radiologists assessed anatomical reproduction, noise, artefacts, and diagnostic quality using ViewDEX. Descriptive statistics and visual grading characteristics analysis were used. In general, DLIR-H scored higher than ASiR-V and FBP. While maintaining image quality, DLIR-H allowed dose reduction compared to FBP. All methods were deemed acceptable for diagnosing pulmonary nodules, fibrosis, and peribronchial pathology. The results indicate that DLIR-H improves image quality in comparison to FBP and ASiR-V and may enable radiation dose reduction while maintaining clinical image quality.