Can deep learning denoising (DLD) significantly reduce radiation dosing without affecting diagnostic quality for head computed tomography (CT) exams?
In a retrospective study, recently published in Academic Radiology, researchers compared DLD and iterative reconstruction (IR) for full-dose (100% mAs) and low-dose simulated head CT scans in 100 cases (mean patient age of 58) involving neuroradiological trauma. Four neuroradiologists provided subjective and objective assessments of image quality, according to the study.
In contrast to the mean pooled subjective ratings for image quality, diagnostic confidence, contrast and sharpness, researchers noted consistently higher ratings for lower radiation dosing with DLD in comparison to IR.
For overall image quality, the study authors noted a mean – 0.720 subjective rating for IR at 25% mAs and a -0.010 rating for DLD at 25% mAs, which was comparable to the -0.008 subjective image quality for IR at 100% mAs. Similarly, when evaluating image sharpness, IR at 25% mAs had a – 0.715 rating while DLD at 25% mAs had a rating of – 0.003, which was similar to the rating for 100% mAs IR (- 0.005). Pooled subjective ratings for DLD at 100% mAs for image quality and sharpness were 0.710 and 0.715, according to the study authors.
“The subjective image quality assessment indicated a reduction of image quality at lower radiation doses when using IR2, while DLD produced high-quality images,” wrote lead study author Georg Gohla, M.D., who is affiliated with the Department of Diagnostic and Interventional Neuroradiology at Eberhard Karls-University Tuebingen in Tuebingen, Germany, and colleagues.
When measuring objective parameters for the comparison of DLD and IR, the study authors noted comparable CT mean attenuation for white matter assessment between IR and DLD at 100% and 25% mAs dosing. The researchers found reduced CT mean attenuation for cerebrospinal fluid (CSF) assessment for 25% mAs DLD (9.60 HU) in comparison to 100% mAs for IR and DLD (10.20 HU) and 25% mAs IR (10.10 HU).
Three Key Takeaways
1. Improved image quality with lower doses. Deep learning denoising (DLD) at 25% mAs produced image quality comparable to full-dose (100% mAs) images with iterative reconstruction (IR).
2. Noise reduction and contrast-to-noise ratio (CNR). DLD demonstrated lower noise levels and higher CNR compared to IR at 100% mAs, suggesting DLD provides superior image clarity even at reduced radiation doses.
3. Potential for radiation dose reduction. The study found that DLD allows for a 75% reduction in radiation dose while maintaining diagnostic image quality equivalent to the full-dose IR, supporting the potential for safer, lower-dose CT imaging in clinical practice.
The study authors also observed that 25% mAs DLD, in comparison to 100% mAs IR, offered reduced noise (8.91 vs. 9.69) and a higher contrast-to-noise ratio (CNR) (2.58 vs. 2.30).
“Our study confirms previous findings that detected lower image noise and higher CNR of deep learning-based than (iterative reconstruction-based) mage series,” pointed out Gohla and colleagues. “ … In our study, images with a 75% reduced dose and denoising technique were equivalent to the 100% original dose images.”
(Editor’s note: For related content, see “AI-Based Denoising for Neck CT May Facilitate Reductions in Radiation Dosing,” “What a Prospective CT Study Reveals About Adjunctive AI for Triage of Intracranial Hemorrhages” and “Head CT for Acute Stroke Patients: Study Shows 22 Percent Improvement for 30-Minute Turnaround Times.”)
Beyond the inherent limitations of a single-center retrospective study, the authors acknowledged the small sample size of 100 adults and the exclusion of non-traumatic neuroradiological emergencies. They also noted diverse diagnoses and multiple pathologies within the reviewed cases of traumatic neuroradiologic emergencies.
