Self-supervised learning for CT deconvolution

Sudhakar, P.; Langoju, Rajesh; Agrawal, Utkarsh; Patil, Bhushan D.; Narayanan, Ajay; Chaugule, Vinay; Amilneni, Vinod; Cheerankal, Paul; Das, Bipul

doi:10.1117/12.2581269

Cited by 1 publication

(1 citation statement)

References 0 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…There are additional reports of using deconvolution-type processing to improve image quality in CT. [33][34][35] More recently, deeplearning-based methods have also been used to improve image sharpness. 36 Apart from deconvolution, partial volume corrections can be applied on calcifications, wherein the assumption is that voxels at an interface are "averaging" a mixture of a calcification and soft tissue. Recently, we developed a method for partial volume correction, 37 and Šprem et al 38 reported a similar approach.…”

Section: Introductionmentioning

confidence: 99%

Improved bias and reproducibility of coronary artery calcification features using deconvolution

Song

Hoori

et al. 2023

J. Med. Imag.

View full text Add to dashboard Cite

Purpose: Our long-range goal is to improve whole-heart CT calcium scores by extracting quantitative features from individual calcifications. Here, we perform deconvolution to improve bias/reproducibility of small calcification assessments, which can be degraded at the normal CT calcium score image resolution.Approach: We analyzed features of individual calcifications on repeated standard (2.5 mm) and thin (1.25 mm) slice scans from QRM-Cardio phantom, cadaver hearts, and CARDIA study participants. Preprocessing to improve the resolution involved of Lucy-Richardson deconvolution with a measured point spread function (PSF) or three-dimensional blind deconvolution in which the PSF was iteratively optimized on high detail structures such as calcifications in images.Results: Using QRM with inserts having known mg-calcium, we determined that both blind and conventional deconvolution improved mass measurements nearly equally well on standard images. Further, deconvolved thin images gave an excellent recovery of actual mass scores, suggesting that such processing could be our gold standard. For CARDIA images, blind deconvolution greatly improved results on standard slices. Bias across 33 calcifications (without, with deconvolution) was (23%, 9%), (18%, 1%), and (−19%, −1%) for Agatston, volume, and mass scores, respectively. Reproducibility was (0.13, 0.10), (0.12, 0.08), and (0.11, 0.06), respectively. Mass scores were more reproducible than Agatston scores or volume scores. For many other calcification features, blind deconvolution improved reproducibility in 21 out of 24 features. Cadaver images showed similar improvements in bias/reproducibility and slightly better results with a measured PSF.Conclusions: Deconvolution improves bias and reproducibility of multiple features extracted from individual calcifications in CT calcium score exams. Blind deconvolution is useful for improving feature assessments of coronary calcification in archived datasets.

show abstract