Deep learning image reconstruction technique for improving image quality and radiation dose reduction compared to iterative reconstruction technique in non-contrast CT head imaging.
Authors
Affiliations (3)
Affiliations (3)
- Department of Medical Imaging Technology, Manipal College of Health Professions, Manipal Academy of Higher Education, Manipal, India.
- Department of Radiodiagnosis and Imaging, Kasturba Medical College, Manipal Academy of Higher Education, Manipal, India.
- Philips Research and Development, Philips Innovation Campus, Yelahanka, Karnataka, India.
Abstract
Non-contrast CT (NCCT) head is a standard imaging modality for evaluating central nervous system pathologies. Repeated CT brain examinations are associated with a cumulative risk of cancer development. Deep learning image reconstruction (DLIR) techniques help improve image quality (IQ) while reducing radiation dose (RD) compared with iterative reconstruction techniques (iDose<sup>4</sup>). Therefore, our study compared IQ parameters among low-dose (LD) CT with DLIR (Precise Image), LD CT with iterative reconstruction (iDose<sup>4</sup>), and standard-dose (SD) CT with iDose<sup>4</sup> and further evaluated radiation dose reduction between SD-iDose<sup>4</sup> and LD-DLIR in NCCT head imaging. Group A (SD with iDose<sup>4</sup>) and Group B (LD with DLIR; Precise Image) each included 96 patients. All NCCT brain scans were performed using a 128-slice incisive CT scanner (Philips Healthcare Systems) with iDose<sup>4</sup> and DLIR. Qualitative and quantitative IQ analyses and radiation dose indices were compared between the groups. Lesion conspicuity was also assessed between LD-DLIR and LD-iDose<sup>4</sup>. There was a significant reduction in RD (effective dose: 1.01 vs. 2.4 mSv; <i>p</i><0.05) with LD-DLIR. LD-DLIR demonstrated superior IQ compared with both SD-iDose<sup>4</sup> and LD-iDose<sup>4</sup>. Subjectively, gray-white matter differentiation improved from scores of 3 (SD-iDose<sup>4</sup>) and 2.5 (LD-iDose<sup>4</sup>) to 4.5 (LD-DLIR), while overall image quality and subjective image noise increased from 3-3.5 (SD-iDose<sup>4</sup>) and 2.5-3 (LD-iDose<sup>4</sup>) to 4-4.5 with LD-DLIR (<i>p</i> < 0.05). Image noise was lowest with LD-DLIR in gray matter thalamus (GMT: 2.30-2.31) compared with SD-iDose<sup>4</sup> (4.80-4.85) and LD-iDose<sup>4</sup> (6.15-6.25); similar results were seen in the white matter posterior limb of the internal capsule (WMPIC: 1.98-2.21 vs. 4.36-4.38 vs. 5.19-5.32) and adjacent cortical gray matter (ACGM: 2.17-2.19 vs. 4.14-4.22 vs. 5.04-5.08). SNR was highest with LD-DLIR [GMT: 16.29-16.36 vs. 7.32-7.34 (SD-iDose<sup>4</sup>) and 5.61-5.75 (LD-iDose<sup>4</sup>); WMPIC: 14.69-14.75 vs. 6.49-6.51 vs. 5.18-5.26], and contrast-to-noise ratio (CNR) was also markedly improved [GMT-WMPIC: 2.59-2.61 (LD-DLIR) vs. 0.98-0.99 (SD-iDose<sup>4</sup>) and 0.82-0.83 (LD-iDose<sup>4</sup>); ACGM-FWM: 2.15-2.29 vs. 0.93-0.95 vs. 0.85-0.86]. Lesion conspicuity was superior with LD-DLIR for all brain lesions compared with LD-iDose<sup>4</sup>. Our study demonstrated that LD NCCT head imaging with DLIR provides superior noise reduction and improved GWMD, SNR, and CNR compared with SD and LD with iDose<sup>4</sup> and supports the clinical utilization of DLIR.