Feasibility of AI-Denoised Ultra-Low-Dose CT for Detection of Acute Pelvic and Hip Fractures: A Pilot Multi-Reader Study.
Authors
Affiliations (5)
Affiliations (5)
- Department of Medicine, University of Massachusetts Chan Medical School, Worcester, 01655 MA, USA.
- Department of Radiology, University of Massachusetts Chan Medical School, Worcester, 01655 MA, USA.
- Department of Radiology, Massachusetts General Hospital, Boston, 02114 MA, USA.
- Department of Radiology, Vanderbilt University, Nashville, 37235 TN, USA.
- VRad, Minneapolis, 55435 MN, USA.
Abstract
<b>Objective:</b> To evaluate the feasibility of simulated ultra-low-dose CT (ULD-CT) enhanced with deep learning denoising (DLD) for detecting pelvic and hip fractures. <b>Method:</b> This Institutional Review Board-approved retrospective study was performed at a multisite, single academic institution between October 2020 and May 2021, including 30 full-dose CT (CTfull) exams of acute pelvic and hip trauma (10 normal, 10 "easy," 10 "hard" cases by fracture conspicuity). Simulated low-dose CT images at 10% (CT10sim) and 5% (CT5sim) radiation dose were generated using a noise model-based algorithm and reconstructed from full-dose CT datasets. A commercial deep learning denoising algorithm enhanced these images, resulting in denoised low-dose CTs (CT10dld and CT5dld) for evaluation. The 90 resulting CT cases were reviewed independently for diagnostic accuracy, confidence, sufficiency, and imaging quality. <b>Result:</b> Diagnostic accuracy was similar across CT modalities overall (CTfull 80.8% [72.9, 86.9]; CT10dld 78.3% [70.1, 84.8]; CT5dld 77.5% [69.2, 84.1]) and within the easy and normal strata; in hard cases, accuracy was numerically lower for CT10dld (45.0%) and CT5dld (42.5%) than for CTfull (57.5%). Diagnostic sufficiency was similarly high across modalities (CTfull 91.7%, CT10dld 87.5%, CT5dld 88.3%). CT full was rated highest in perceived image quality across all strata (mean 4.21 ± 0.83 overall vs. 3.47 ± 0.97 for CT10dld and 3.23 ± 0.98 for CT5dld) and in diagnostic confidence (mean 4.61 ± 0.63 vs. 4.45 ± 0.71 and 4.41 ± 0.72, respectively), though neither translated into a corresponding drop in diagnostic accuracy overall. Inter-reader agreement was highest between MSK radiologists, reaching substantial agreement on CT10dld (κ = 0.760) and CT5dld (κ = 0.672). <b>Conclusions:</b> DLD-enhanced ULD-CT showed diagnostic accuracy and sufficiency similar to full-dose CT for pelvic and hip fracture detection, at a fraction of the radiation dose, though full-dose CT was rated higher in perceived image quality and diagnostic confidence and outperformed denoised low-dose CT numerically in hard cases. Full-dose CT remains the preferred modality for complex and high-energy trauma, while denoised low-dose CT modalities represent a promising option for straightforward presentations where radiation minimization is a priority.