Implementation of AI image reconstruction in CT—how is it validated and what dose reductions can be achieved

Abstract

CT reconstruction has undergone a substantial change over the last decade with the introduction of iterative reconstruction (IR) and now with deep learning reconstruction (DLR). In this review, DLR will be compared to IR and filtered back-projection (FBP) reconstructions. Comparisons will be made using image quality metrics such as noise power spectrum, contrast-dependent task-based transfer function, and non-prewhitening filter detectability index (d_NPW'). Discussion on how DLR has impacted CT image quality, low-contrast detectability, and diagnostic confidence will be provided. DLR has shown the ability to improve in areas that IR is lacking, namely: noise magnitude reduction does not alter noise texture to the degree that IR did, and the noise texture found in DLR is more aligned with noise texture of an FBP reconstruction. Additionally, the dose reduction potential for DLR is shown to be greater than IR. For IR, the consensus was dose reduction should be limited to no more than 15–30% to preserve low-contrast detectability. For DLR, initial phantom and patient observer studies have shown acceptable dose reduction between 44 and 83% for both low- and high-contrast object detectability tasks. Ultimately, DLR is able to be used for CT reconstruction in place of IR, making it an easy “turnkey” upgrade for CT reconstruction. DLR for CT is actively being improved as more vendor options are being developed and current DLR options are being enhanced with second generation algorithms being released. DLR is still in its developmental early stages, but is shown to be a promising future for CT reconstruction.

Introduction

CT reconstruction continues to evolve with improvements in computational technology. The long-used filtered back-projection (FBP) reconstruction algorithm has served the CT medical community well since its introduction in the mid-1970s. ¹ FBP is mathematically straightforward and computationally quick. One of the major benefits of FBP is the “filter” portion of the reconstruction algorithm. Back projected-only images, alone, would be blurry and non-diagnostic. With the introduction of a mathematical filtering function to the reconstruction process, the blurriness of the image is reduce in favor of object-boundary sharpening. ² However, those same filters that removed blur and sharpened object boundaries also enhanced image noise in the FBP images. Further filtering of the image using specialized reconstruction kernels were introduced to help mitigate image noise. ²

There are two components to image noise: electronic noise and quantum (or statistical) noise. Electronic noise manifests as an underlying, low-amplitude signal generated by the detector electronics themselves; electronic noise is present even when the detectors are not exposed by x-rays. Electronic noise is only prevalent in reconstructed images when a CT operates at very low X-ray levels. Technological advancements of detector construction has helped to minimize the effect of electronic noise in a reconstructed CT image.

In CT, quantum noise usually dominates; thus, when we refer to image noise, we are referring to quantum noise alone. Quantum noise arises from the random nature of X-ray interactions, governed by Poissonian counting statistics. The randomness, or fluctuations, of the X-ray signal in the detector causes uncertainty in the reconstruction algorithm. The uncertainty manifests as variations in the calculated CT number; thus, in uniformly attenuating regions of the body where CT numbers should be the same, neighboring pixels will have slightly different CT numbers leading to pixel-to-pixel grayscale variation or streaking artifacts. For FBP reconstruction, quantum noise can be reduced by increasing the number of X-rays, i.e. radiation dose, on the CT detector; thus, reducing the relative amount of fluctuation within the total readout signal. Therefore, removing image noise in FBP imaging is related to the amount of radiation output, and conversely, the radiation dose as shown in Eq.1:

$$Noise=\frac{1}{\sqrt{dose}}$$ (1)

Hence, dose reduction using low tube current time product (i.e. mAs) or low tube potential (i.e. kV) acquisition techniques is limited using FBP because noise will increase as those technique factors decrease.

To overcome the limitation of image noise in the reconstruction process, iterative reconstruction (IR) methodologies were explored in the mid-2000s. ^3,4 The very first CT reconstruction method was actually iterative in nature; it was called algebraic reconstruction technique (ART), ⁵ but due to limited computational power in the 1970s, ART was replaced with FBP. Even in the mid- to late-2000s with all the advancements in computational power at that time, fully iterative, model-based iterative reconstruction (MBIR) algorithms were too computationally intense to be used clinically ^4,6 ; during the initial development of MBIR algorithms, reconstruction times took hours to days to produce a series of CT images. The reason MBIR image reconstruction initially took so long was that MBIR accounts for both the image noise properties and geometry of the CT system (i.e. the focal spot and detector sizes, distance between X-ray tube focal spot and detector, etc.). Though a slower reconstruction process, by including the physics of the imaging chain into the reconstruction process, MBIR has been shown to reduce image noise, ⁷ enhance object boundaries in the body, ² and reduce artifacts. ⁸

As a compromise, statistical-based iterative reconstruction (SBIR) algorithms were introduced in 2008 ³ as advanced reconstruction models that mixed the speed of FBP with an iterative approach that was able to reduce quantum noise in the FBP image. SBIR algorithms were highly effective at reducing image noise only, and widely adopted across all CT manufacturers. As computational speeds increased, and SBIR algorithms became more integrated into routine clinical care, hybrid-IR algorithms were introduced that blended noise reduction capabilities from SBIR with some improvements seen using MBIR (such as enhanced image resolution and artifact reduction). The hybrid-IR algorithm reconstruction times were on the order of a SBIR algorithm, ⁶ but provided some enhancements that were desirable from slower MBIR algorithms. ^4,6,9

However, applying IR led to reconstructed images appearing smoother and less pleasing to the eye. ¹⁰ For several decades, radiologists became accustomed to the sharper object boundaries and noise texture depicted in FBP images, IR was shown to adversely affect noise texture and smooth object boundaries. IR introduced non-linear effects such that image noise was spatially dependent, and spatial resolution was contrast dependent; thus, low-contrast and low-frequency objects in the body were more negatively impacted compared to high-contrast, high-frequency structures. ^11,12 To minimize the change in appearance of CT images using IR, IR algorithms are used but are generally used at lower setting levels; the lower setting levels of IR have been shown to minimize object boundary softening and maintain noise texture appearances as previously seen in FBP-only reconstructions, ^6,13–18 but commensurately, do not remove as much image noise.

Overview of deep learning reconstruction application to CT reconstruction

CT reconstruction using artificial intelligence, referred to as deep learning reconstruction (DLR), has introduced a new way to improve image quality without the penalty of increasing patient radiation dose or reconstruction time. Commercially available DLR algorithms use a deep convolutional neural network (DCNN) to recognize image noise patterns and remove those noise patterns from the raw data or from the reconstructed CT image. Critical to the success of any DCNN is the training data sets employed to teach the algorithm to differentiate between noise and anatomical structure. Each DLR developer has approached DCNN training differently, ^17,19,20 but the quality and type of data used for training directly impacts the final reconstructed DLR image. In all cases, DCNN algorithms are trained using pair-wise data of low noise images (defined as ground truth) and noisy images. Through the training process the selection of hyperparameters and node weights are tuned to maximize DLR consistency and accuracy. The benefit of a pre-trained DCNN is: consistent results and fast-throughput reconstruction, which benefits are in contrast to MBIR algorithms that tune the model in real-time during reconstruction and thus requires greater reconstruction time. ¹⁷ Additionally, DLR algorithm reconstruction times are nearly equivalent to those of SBIR due to the innovation and deployment of graphic processing unit (GPU) enabled computers.

There are many different approaches to the use of DLR that have been explored in the scientific literature, but the two primary approaches available to purchase are: (1) a vendor-independent approach that operates the DLR algorithm in image-space and (2) a vendor-specific approach that operates in image-space combined with projection-space, since the DLR algorithm has access to the raw, sinogram-data on the CT scanner. The first approach is primarily offered by third-party solutions that sit between the CT scanner and the PACS station. The CT scan is performed and the data are initially reconstructed using the reconstruction algorithm of choice on that CT scanner. The CT scans are then sent to an independent server that will function to clean up the CT reconstructions using vendor-independent software. ^21,22 Because these DLR algorithms do not have access to the raw scan data, their primary function is to denoise the image. The second, vendor-specific approach, likewise functions primarily to reduce image noise, but has been shown to also improve object-edge definition and improve spatial resolution due to its access to projection-space data. ^19,23–28 All current versions of DLR come with multiple strengths or implementation level options. Some DLR options, such as Canon’s AiCE (Canon Medical Systems, Otawara, Japan) are available for use on all body parts, whereas GE’s TrueFidelity (GE Healthcare, Waukesha, WI) is only applicable to image reconstruction applied with the standard reconstruction kernel, which is typically only applied for soft tissue organ reconstruction.

Validation and quality assurance of DLR for CT

CT DLR algorithms have been shown to improve image quality by reducing image noise, preserve (and in some cases improve) object edge detection, and enhance diagnostic confidence for various diagnostic tasks. ^19–28 Since noise in a CT image is often the limiting factor when low-contrast detectability is the primary diagnostic task, e.g. the visibility of focal liver lesions and infection can be suppressed with higher levels of imaging noise, the validation and characterization of image noise properties for DLR applications is critical.

DLR, like IR, exhibits non-linear noise parameters such that classic image quality metrics as pixel noise, contrast-to-noise ratio, and signal-to-noise ratio, on their own, do not fully convey the impact of DLR on image noise. ¹⁷ The use of Fourier-based image quality metrics such as noise power spectrum (NPS) and contrast-dependent task-based modulation transfer function (TTF) have been widely used when characterizing image noise, contrast, and resolution in CT, when applied to FBP, SBIR, and MBIR. However, the use of NPS and TTF for CT is limited. The definition of NPS assumes wide-sense stationary noise and TTF assumes linear, shift-invariant systems. ¹² In all cases: FBP, SBIR, MBIR, and DLR violates one or more assumptions for use of Fourier-based image quality metrics. The calculation of NPS ²⁹ and TTF ¹¹ must be performed within the bounds of the calculation limitations, and the interpretation of results from a study utilizing NPS and TTF can only be applied within the context of the defined limitations, especially when applied to DLR. ³⁰ The report by Vaishnav et al ¹² advocates for an alternative, namely: to use task-based observer studies (either mathematical or human) as the gold standard by which diagnostic efficacy, image quality, and dose reduction can be assessed for DLR CT. In the work by Solomon and Samei, ³¹ comparison between a human detection study, considered the gold-standard, and several mathematical observer methodologies, calculated using Fourier-based image metrics, demonstrated strong correlation, based on their assumptions of quasi-linearity and local noise stationarity; the results from this study demonstrate the way in which Fourier-based metric calculations may still be used for CT analysis.

Since FBP images have been held as the golden standard for preferred noise texture and image object boundary sharpness, in this review, comparisons of hybrid-IR and DLR image quality metrics were made with respect to FBP reconstructed images using an American College of Radiology (ACR) CT accreditation phantom. ³² NPS plots for hybrid-IR and DLR, for two vendor-specific solutions, were compared to FBP NPS plots, [Figure 1]. To calculate NPS, local noise stationarity was assumed within regions of interest measuring 92 pix² using a mean subtraction method to remove DC offset. To quantify the noise reduction potential, across all frequencies (i.e. noise textures), the noise magnitude ratio (NMR) ³³ was calculated, as shown in (Eq. 2):

Figure 1.

Noise power spectrum plots were derived using an ACR CT phantom (scan technique factors were held constant for all vendor-specific image acquisitions) and demonatrate noise variance reduction along the ordinate axis as a function of noise frequency (i.e. noise texture) for FBP, hybrid-IR, and DLR images along the abcissa. (a) Canon’s AiCE is shown at three strengths of DLR implementation along side a FBP and AIDR3D (i.e. hybrid-IR); (b) plot is simplified to only show AiCE at three strengths. (c) GE’s TF is shown at three strengths of DLR implementation alongside a FBP and 40% implementation of ASiR-V (i.e. hybrid-IR); (d) plot is simplified to only show TF at three strengths. ACR, American College of Radiology; DLR, deep learning reconstruction; FBP, filtered back-projection; IR, iterative reconstruction; TF, TrueFidelity.

$$NMR=\sqrt{\frac{\int {NPS}_{IRorDLR}\left(f\right)df}{\int {NPS}_{FBP}\left(f\right)df}}.$$ (2)

NMR values < unity indicate greater noise reduction potential for hybrid-IR or DLR, as compared to FBP, and NMR values equal to unity indicate no change in noise magnitude. Additionally, to assess the impact hybrid-IR and DLR algorithms have on changing image noise texture, the center frequency ratio (CFR) ³³ was calculated as shown in (Eq. 3):

$$CFR=\frac{{CF}_{IRorDLR}}{{CF}_{FBP}}$$ (3)

CFR defines the magnitude to which noise texture is altered, with respect to FBP reconstruction by assessing the magnitude of change of the center frequency (CF) of the NPS curves. CFR values equal to unity represent no shift in CF and by extension, noise texture. CFR values < unity represent images with lower-frequency, coarser, image noise structure, and CFR values > unity represent images with higher-frequency, finer, image noise structure.

Results from this NPS analysis demonstrate CFR values are near unity for all levels of DLR, which demonstrates that DLR algorithms can better remove noise at all frequencies from the image thereby maintaining noise texture in the reconstructed image similar to that of FBP [Table 1]. For one vendor-specific DLR implementation, the CFR values are greater than unity. For this specific vendor, the DLR algorithm was trained using MBIR images that had a strong spatial resolution enhancement feature, as seen in Figure 2(a). These CFR results stand in contrast to the well-documented results that demonstrate biased reductions of higher frequency noise by IR algorithms which led to softer, unpleasant looking reconstructed images.

Table 1.

Noise power spectrum plots, [Figure 1], were quantitatively analyzed for FBP, hybrid-IR, and three strengths of DLR data derived from images acquired using an ACR CT phantom

		FBP	Hybrid-IR	DLR-Low	DLR-Med	DLR-High
Canon	CF (lp/cm)	0.28	0.26	0.38	0.36	0.26
	NMR	1	0.69	0.51	0.43	0.28
	CFR	1	0.93	1.36	1.27	0.93
GE	CF (lp/cm)	0.31	0.23	0.29	0.28	0.26
	NMR	1	0.72	0.86	0.77	0.67
	CFR	1	0.75	0.93	0.89	0.84

ACR, American College of Radiology; CF, center frequency; CFR, center frequency ratio ; DLR, deep learning reconstruction; FBP, filtered back-projection; NMR, noise magnitude ratio.

Two DLR algorithms were compared: AiCE (Canon Medical Systems) and TrueFidelity (GE Healthcare).

Figure 2.

TTF plots were derived using an ACR CT phantom (scan technique factors were held constant for all vendor-specific image acquisitions; TTF was calculated using the high-contrast sensitometry insert, i.e. ~900 HU) and demonatrate image contrast along the ordinate axis as a function of object size (i.e. spatial frequency) along the abscissa for FBP, hybrid-IR, and DLR images. (a) Canon’s AiCE is shown at three strengths of DLR implementation along side a FBP and AIDR3D (i.e. hybrid-IR); (b) GE’s TF is shown at three strengths of DLR implementation alongside a FBP and 40% implementation of ASiR-V (i.e. hybrid-IR). Horizontal dashed lines are provided as common points of reference at the 50% and 10% TTF. ACR, American College of Radiology; DLR, deep learning reconstruction; FBP, filtered back-projection; IR, iterative reconstruction; TF, TrueFidelity; TTF, task-based modulation transfer function.

The DLR results generally demonstrated reduced NMR values [Table 1], compared to hybrid-IR NMR values. Thus, the resulting image retains more of the noise texture characteristics (as described by CFR values near unity) but at lower noise levels then seen in current-day implementations of hybrid-IR algorithms (as described by lower NMR values). Similar NPS analysis results, as presented in this work for DLR, have been reported. ^{17,19,23,24,27,30,33–35}

Differences in vendor-specific spatial resolution recovery are demonstrated when comparing plots of TTF for FBP, hybrid-IR, and DLR [Figure 2]. The behavior of FBP and hybrid-IR TTF curves shown in [Figure 2] are typical of previous reports. ^36,37 The difference in DLR TTF responses between vendor implementations is attributed to different methods for DLR training undertaken by each vendor, respectively. As previously described, Canon’s AiCE, [Figure 2(a)] was primarily trained using a MBIR algorithm that both reduced image noise and enhanced object sharpness. ^24,34,35,38 GE’s TrueFidelity, [Figure 2(b)], was primarily trained using FBP only images, hence the DLR TTF curves are nearly identical to the FBP TTF curve. ^25,30

Recently, Canon released a second generation of DLR called Precise IQ-Engine (PIQE). ^17,39 The PIQE algorithm’s training incorporated unique data acquired on a high-resolution CT scanner ⁴⁰ (CT scanner detectors were 0.25 × 0.25 mm compared to traditional 0.5 × 0.5 mm detector elements used in the training of AiCE). The incorporation of high-resolution data in the training algorithm has resulted in a new, enhanced resolution recovery DLR that can better remove image noise while preserving object boundaries; thus, rendering traditional CT images acquired on a 0.5 × 0.5 mm detector system to have spatial resolution-equivalent images as if the images were acquired on a high-resolution 0.25 × 0.25 mm detector system. PIQE is currently available for cardiac applications only.

To quantify any improvement in diagnostic accuracy or diagnostic confidence when using DLR vs hybrid-IR or FBP, traditionally, reader observer studies are used to calculate a receiver operating characteristic (ROC). ROC analysis of DLR reconstructed clinical images, based on radiologist ratings, have demonstrated improvements in image noise, ^{23,28,38,41–46} edge sharpness preservation, ^38,46 reading time reduction, ^47,48 and artifact reduction. ^38,43 Similarly, studies have claimed that clinical-finding detection rates (e.g. lesions, fractures, coronary artery disease, etc.) using DLR was equal to or better than hybrid-IR ^42,47,48 ; note, when different levels of DLR were compared, some studies indicated that low-contrast lesion detection was ideal at the medium DLR setting, and that at the highest DLR setting, the low-contrast lesions became less conspicuous (as compared to lower DLR settings), but were still observed. ^44,45

Often it is difficult to conceptualize how NPS and TTF can independently be used as figures of merit to characterize image quality and their impact on diagnostic accuracy and diagnostic confidence. One method is to incorporate NPS and TTF into one metric called the non-prewhitening filter detectability index (d_NPW'). ²⁴ The metric d_NPW' also incorporates other facets of image quality such as: (1) how the size, shape, and contrast of an object ^31,49 is impacted by the reconstruction algorithm; (2) how the human visual system impacts the observers ability to detect the object by accounting for sensory noise in human visual system ^50,51 ; (3) and how the eye functions at different distances (from a displayed image) and object frequencies (i.e. low-contrast, soft edges vs high-frequencies, strong edges). ^31,50 Prior studies have validated the accuracy of d_NPW' as a mathematical observer model, when compared with radiologists, as a surrogate for object detectability. ^31,50–54 Using d_NPW', allows studies to predict diagnostic confidence as a function of many different CT acquisition techniques and reconstruction parameters. Previous studies using d_NPW' have demonstrated superior noise reduction, edge sharpness preservation, and improved clinical-finding detection rates for DLR compared to IR and FBP. ^24,35,55

Dose reduction potential using DLR

The technological impetus to move away from FBP was largely due to increasing worry around CT radiation dose to the medical population at the turn of last century. ^56,57 With the developing urgency for better radiation dose control using CT, various approaches to radiation dose reduction have been implemented to great success: tube current modulation, ^58–61 use of lower tube potential (i.e. kV), ^60–63 the development of weight- or size-based protocols, ^60,61,64 and the development of appropriateness criteria. ^65,66 Each approach has been successful in reducing CT radiation dose, especially in the pediatric population, ⁶⁷ but the continued reliance on FBP was the limiting factor for substantial dose reduction in CT since low technique factors (such as low mAs and or low kV) resulted in signal-starved or very noisy images that may have masked diagnostic findings.

The development of IR was the beginning of a process to decouple image noise in the reconstructed image with the radiation dose. With the introduction of IR, image noise could be algorithmically removed irrespective of the mAs or kV used to acquire the image. Similarly, with the introduction of DLR, DLR offers the same potential dose reduction as has been demonstrated with the introduction of IR. To explore the potential for dose reduction, an ACR CT phantom was scanned using two vendor-specific DLR algorithms at the medium setting and at decrementing mAs values starting at 300 mAs down to 10 mAs. The NPS was derived from each set of images and CFR and NMR values were calculated for each dose reduced spectrum and plotted as a function of mAs. The CFR was analyzed as a ratio of the NPS central frequency at the high dose, 300 mAs, compared to the dose reduced DLR NPS central frequencies; thus, as radiation was reduced and image noise increased the CFR ratio described the effect of noise texture in the reconstructed image. CFR values < unity represent images with lower-frequency, coarser, image noise structure, and CFR values > unity represent images with higher-frequency, finer, image noise structure. The NMR was analyzed as a measure of increased noise magnitude at decreasing radiation levels after DLR was applied to remove image noise.

The results for the two vendor-specific DLR algorithms demonstrates unique differences in noise texture at extremely low dose levels. One vendor’s DLR algorithm is biased to remove high-frequency noise content, leading to a sharp CFR reduction at low doses, whereas the other vendor’s DLR algorithm is biased to remove low-frequency noise content, leading to a slight increase in finer textured noise at low doses, [Figure 3]. Results for the NMR analysis demonstrated different levels of noise control using DLR, [Figure 4]. As radiation dose decreased, noise increased to a maximum of 264% for AiCE and 972% for TrueFidelity, as compared to the noise magnitude in the control image acquired at 300 mAs. The AiCE DLR algorithm maintained noise texture to within 2% down to a dose reduction of 70% (1–90 mAs/300 mAs) and a NMR increase of 57%, whereas TrueFidelity maintained noise texture to within 7% and a NMR increase of 315% down to a dose reduction of 70%, [Table 2]. In comparison, several studies used phantoms to quantitatively compare DLR with SBIR and or FBP. Racine et al ⁶⁸ compared NMR and CFR values for DLR (i.e. TrueFidelity) to FBP and demonstrated similar noise texture preservation to within 5.5% and NMR increase of 38% at 50% dose reduction. Using an anthropomorphic physical phantom, Lee et al ⁶⁹ conducted a reader observer study including five radiologists, and demonstrated radiation dose reduction potential on the order of 65.5–68.1% for a DLR algorithm (i.e. TrueFidelity), as compared to FBP and SBIR algorithms; in that study, the observers rated the quality of organ tissue interfaces, hepatic vascularity, and overall image quality satisfactorily. In an objective and reader-based subjective phantom study, Park et al ⁷⁰ demonstrated improved image noise and reduced artifacts for a DLR (i.e. TrueFidelity) to that of SBIR and equivalent image quality down to 70% dose reduced images. Dose reduction estimates derived from phantoms alone are limiting, namely: phantoms are finite in size and may not be generalizable over a wide range of body habitus, and the attenuation characteristics of a phantom are different than humans thus leading to different noise absolute levels and patterns.

Figure 3.

CFR was calculated at decrementing radiation output levels (i.e., mAs) for Canon’s AiCE and GE’s TrueFidelity. The fitting function for AiCE took the form ${\displaystyle a\left(1-e^{-bx}\right)+c}$ with coefficient values: a = 0.6, b = 0.03, and c = 0.41. The SSR was 0.001. The fitting function for TrueFidelity took the form ${\displaystyle a\left(x\right)+b}$ with coefficient values: a = −0.0003 and b = 1.11. The SSR was 0.003. CFR, central frequency ratio; SSR, sum of squared residual.

Figure 4.

NMR was calculated at decrementing radiation output levels (i.e., mAs) for Canon’s AiCE and GE’s TrueFidelity. The fitting function for both AiCE and TrueFidelity took the form $\frac{1}{\left(\frac{1}{a}+bx\right)}+c$ . The coefficient values for AiCE were: a = 2.31, b = 0.01, and c = 0.66 with a SSR of 0.003. The coefficient values forTrueFidelity were: a = 14.97, b = 0.01, and c = 1.12 with a SSR of 0.82. NMR, noise magnitude ratio; SSR, sum of squared residual.

Table 2.

NMR and CFR were calculated from images acquired using an ACR CT phantom imaged at decrementing radiation output (i.e. mAs) for two vendor-specific DLR algorithms

	AiCE		TrueFidelity
mAs	CFR	NMR	CFR	NMR
10	0.58	2.64	1.08	9.72
30	0.80	2.16	1.12	5.43
50	0.89	1.87	1.10	4.24
70	0.95	1.69	1.10	3.51
90	0.98	1.57	1.10	3.15
110	1.00	1.47	1.07	2.84
130	1.00	1.39	1.05	2.66
150	1.02	1.31	1.04	2.46
170	1.02	1.25	1.08	2.43
190	1.03	1.20	1.05	2.33
300	1.00	1.00	1.00	1.00

ACR, American College of Radiology; CFR, center frequency ratio; DLR, deep learning reconstruction; NMR, noise magnitude ratio.

Several studies used patient (or cadaver) studies to investigate the impact of DLR dose reduction on various diagnostic tasks. Miyata et al ⁷¹ demonstrated dose reduction to ~85% in a reader study including three radiologists when comparing DLR (i.e. TrueFidelity) and FBP; in that study, small vessels and a generic assessment of image quality (namely: noise magnitude, streak artifacts, general appearance) in four cadaver lung CT images was performed. Nam et al ²² collected 100 consecutive patients undergoing chest CT processed with DLR (i.e. TrueFidelity) and abdomen CT processed with SBIR. The upper abdomen images from the ~50% lower dose chest CT processed with DLR were compared to the abdomen images processed with SBIR in a reader observer study. Objective measures of image noise and pooled observer preference both scored higher for DLR compared to SBIR. Tamura et al ⁷² compared 71 patients imaged at 44% reduced dose reconstructed with DLR (i.e. AiCE) to similar patient populations imaged with SBIR; two radiologists rated the DLR images to be superior based on general image quality standards and diagnostic confidence to assess hepatocellular carcinoma lesions > 1 cm.

As has been demonstrated, the use of DLR can be thought of as a dose reduction tool in a similar way as IR was when it was first introduced to the CT community. With the early adoption of IR, many users reported the use of IR for dose reduction purposes in lieu of an image denoising tool, as initially, the algorithm was envisioned. There were many approaches to use IR as a dose reduction tool presented in the scientific literature, the most straightforward approach was to sample image noise in a FBP-reconstructed patient population then using IR, the user could reduce the radiation output (i.e. lowering mAs, kV, and or other acquisition parameters) and compensate for the increased image noise by using IR to reduce the image noise back to the original FBP-reconstructed levels. ⁷³ Reported results described relatively consistent image noise but at reduced radiation dose when implementing IR. ^14,73–76 This simple approach, though popular in the early days of IR failed to address the loss of object boundary sharpness of low-contrast objects at low radiation doses. ⁷⁷ Subsequent studies demonstrated reduced dose CT, using IR for noise control, were not diagnostically equivalent when performing a low-contrast lesion detection task. Following a large reader study, the consensus was dose reduction should be limited to no more than 15–30% to preserve low-contrast detectability. ^77–83 In other words, the non-linear approach to noise-mitigation using IR led to softening of object boundaries and loss of low-contrast detectability.

For DLR to function as a clinical dose reduction technique, the ability to conserve detection accuracy, especially for low-contrast tasks at low doses must be demonstrated. In a study by Singh et al, ⁸⁴ 59 patients receiving chest and abdominopelvic CT examinations underwent a comparative prospective study where all patients were imaged at normal clinical dose followed by an 83% (1–2.1mGy/13 mGy) dose reduced CT; in that study, all abdominal lesions and pulmonary nodules were seen at low dose DLR (i.e. AiCE) as confirmed by normal dose SBIR, MBIR and FBP. In a liver metastases detection study, Lyu et al ⁸⁵ demonstrated non-inferiority down to 50% for all size lesions, but down to 70% reduction for metastasis >1 cm. Zhao et al ⁸⁶ compared RECIST measurement accuracy for DLR between low dose CT (0.07 mSv) and normal dose CT (2.38 mSv) with pulmonary lesion measurements agreeing to better than 2.2% and lymph node measurements agreeing to better than 1.4%. In a high-contrast detection study, Qu et al ⁸⁷ compared arterial stenosis measurements made for lower dose extremity CTA and found them to be equally accurate with better image quality and sharper vessel wall visualization than clinical dose SBIR CT. Noda et al ⁸⁸ evaluated low dose (2.3 mGy) abdominal CT for conspicuity of pancreatic ductal adenocarcinoma in 28 prospective patients; DLR (i.e. TrueFidelity) scored higher than IR when evaluated for diagnostic confidence and conspicuity. Brady et al ²⁴ performed a reader-based retrospective study to asses general image quality and low- and high-contrast object detectability for a pediatric patient population; three pediatric radiologist rated DLR (i.e. AiCE) image quality superior to SBIR, MBIR, and FBP with increased diagnostic confidence at a dose reduction potential of 52%.

Limitations of this review paper: review papers provide an opportunity to synthesize the current state of knowledge surrounding a topic, in this case, DLR for CT image quality and dose reduction potential, and can be thought of as a waymarker to guide future researchers. The results summarized and presented in this work represent the current state of practice of DLR, which results have made strong claims of dose reduction (44–83%) while preserving or improving key image quality features such as: noise texture similar to FBP, object boundary sharpness, artifact reduction, general observer preference, all with similar diagnostic confidence for low- and high-contrast diagnostic tasks. The claims of dose reduction demonstrated in the publications cited herein must be understood and interpreted within the boundary conditions of those studies; extrapolation beyond the methods of those studies should be made cautiously. DLR has been shown to provide positive improvements in image quality; however, the magnitude of dose reduction is still an open question worthy of debate.

Prior studies investigating the implementation of IR should be used to both demonstrate the strengths and weaknesses of current investigative tools used when characterizing DLR in CT, and demonstrate potential danger of misuse of those tools. In the early days of IR, those publications tended to over predict dose reduction potential for IR, especially for low-contrast detectability tasks. Ongoing investigations for the use of DLR as a dose reduction tool need to focus on detection accuracy of different pathologies, object sizes, and object contrasts, all at different radiation dose levels and in terms of different DLR strengths. Additionally, there is a specific need for more studies looking at the impact of DLR on various diagnostic tasks related to pediatrics. Finally, quantitative metrics need to be employed to demonstrate that DLR technology is not masking or creating pathologic features in the reconstructed CT, and ideally, beyond simple detection confidence analysis, investigators should continue to explore the question: does DLR increase detection rate and accuracy and is there a relationship between accuracy and radiation dose level?

Conclusion

CT reconstruction has undergone a substantial change over the last decade with the introduction of IR and now with DLR. IR algorithms were able to weaken the link between image noise and radiation dose, allowing a decade of decreasing radiation dose levels with image noise levels staying largely the same. However, the use of IR led to fundamental changes in how image noise was perceived and the softening of low-contrast object boundaries; both changes were shown to affect diagnostic confidence for low-contrast object detectability leading to a limited implementation of IR. In contrast, DLR algorithms have been shown to better preserve noise texture (i.e. noise texture in DLR is similar to that of FBP) and object boundary sharpness at dose levels currently used in conjunction with IR. The potential for dose reduction using DLR has been shown to be feasible down to 44–83% while maintaining equivalent image quality and diagnostic confidence as compared to IR and FBP CT. Further research in the use of DLR for various different diagnostic tasks, beyond simple noise removal, is still needed. DLR for CT is actively being improved as more vendor options are being developed and current DLR options are being enhanced with second generation algorithms being released. Like IR was a decade ago, DLR for CT is still in its infancy, but has been shown to be promising.