Deep learning versus iterative image reconstruction algorithm for head CT in trauma

There is wide consensus that noncontrast head CT is the initial imaging modality of choice for acute moderate to severe traumatic brain injury. CT is specific and sensitive for detecting intracranial hemorrhage (ICH) [1]. However, certain anatomical regions such as the middle and the posterior fossa are more prone to beam hardening artifacts from adjacent dense skull structures which can obscure subtle traumatic lesions [2‐4]. Furthermore, CT without intravenous contrast agents has intrinsically relatively limited soft tissue contrast resolution [5, 6], e.g., in the brain parenchyma where the difference in CT attenuation between gray and white matter is relatively small. The diagnosis of some intracranial pathologies relies on the detection of the discreet alteration in attenuation that they cause between the gray and white matter [5]. These factors can make the interpretation of noncontrast trauma head CT challenging.

Iterative reconstruction (IR) techniques have demonstrated superior image quality for head CT compared to filtered back projection (FBP) with reductions in artifacts and image noise [7‐10]. One reported disadvantage of IR compared to FBP is that its noise reduction can cause a blotchy or “pixelated” image texture [11‐13] which is more pronounced with higher IR strengths [14‐16]. Furthermore, studies have shown that IR achieves inferior resolution compared to FBP for low-contrast features [17], especially at low doses [18, 19].

Artificial intelligence (AI) has in recent years unlocked new possibilities in medical imaging including new methods for CT image reconstruction where deep learning (DL), a subset of machine learning, has been implemented [20]. DL is based on an artificial neural network inspired by the human brain in the sense that it automatically and unsupervised learns distinctive features from the input data itself which gives DL the ability to estimate highly complex nonlinear relationships. The term “deep” refers to the multilayer networks of artificial neurons [20, 21].

The DL-based image reconstruction (DLIR) algorithm TrueFidelity has recently been introduced by GE Healthcare. It has been trained on ground truth data comprising CT images reconstructed with FBP from a great number of phantom and patient cases including different anatomies, clinical indications, and scanning conditions. Both a low-dose/high-noise and high-dose/low-noise dataset was obtained for each case and the latter was used as the ground truth data. The DLIR algorithm was applied on the low-dose/high-noise datasets with the high-dose/low-noise datasets being the training target. CT image quality experts from GE Healthcare and radiologists supervised the training process. TrueFidelity offers three selectable strength levels (low, medium, and high) [22].

As novel image reconstruction algorithms are put on the market, it is important to conduct an independent assessment of their effect on image quality. Recent studies have shown that DLIR algorithms improve head CT image quality compared to IR [23‐25]. However, none of these studies has evaluated how a DLIR algorithm performs compared to an IR algorithm on noncontrast head CT from a cohort of multitrauma patients presenting to a level 1 trauma unit. Our hypothesis is that DLIR will improve the conspicuity of trauma-related findings by decreasing the beam-hardening artifacts and reducing the image noise level, compared to IR. Hence, the purpose of this study was to evaluate the qualitative and quantitative image quality of noncontrast trauma head CT reconstructed with TrueFidelity with all three DLIR strengths and compare it to the IR algorithm adaptive statistical iterative reconstruction-Veo (ASiR-V), also by GE Healthcare.

This study has been approved by the regional ethics committee (Dnr 2020–04700, 2021–02547).

We compared the image quality between a DL-based (TrueFidelity) and an IR-based (ASiR-V) image reconstruction algorithm for trauma head CT. We have demonstrated that TrueFidelity outperforms ASiR-V regarding both quantitative and qualitative image quality parameters. The image quality increased with higher DLIR strengths. In particular, DLIR-M and DLIR-H had significantly lower image noise levels and higher SNR and CNR compared to ASiR-V. DLIR-M and DLIR-H also received higher reader scores compared to ASiR-V and DLIR-L. Furthermore, the proportion of non-diagnostic CT series was lower for DLIR-M and DLIR-H when compared to ASiR-V, for all readers. The improvement in qualitative image quality provided by DLIR was greater for the second most and the less experienced reader than for the most experienced reader.

These findings are in line with the current literature which shows that DLIR algorithms achieve superior image quality for head CT compared to IR-based algorithms [23‐25, 29]. We confirm the findings from a similar study by Kim et al. [23]. However, they included all patients undergoing a noncontrast head CT, no matter the indication, and excluded pathology. We showed that the superior image quality of TrueFidelity compared to ASiR-V also applies to trauma patients with the potential of improving diagnostic performance regarding ICH. Furthermore, we recruited three readers with different levels of experience, as opposed to the study by Kim et al. in which two experienced neuroradiologists performed the qualitative image quality analysis. Not only were the readers in our study blinded to each other’s evaluations, but also to each other’s identity. Another strength with our study is that each reader received an individual training session before initiating image evaluation. Additionally, although debatable, we consider our approach of presenting the ordinal data (reader scores) as percentages a more appropriate measure than the mean. We also included a larger number of patients in our study.

The study by Sun et al. was performed on a pediatric population and even though trauma cases were included in this study the patient cohort was quite heterogenous and also included patients with convulsion or mental symptoms or cases where emergency doctors requested a CT scan to exclude intracranial pathology [24]. We have focused on the trauma patient subset exclusively and thereby showed that DLIR can successfully be implemented for head CT on trauma patients presenting to a level 1 trauma unit. Additional strengths with our study are that we almost had the double sample size compared to Sun et al.

In a retrospective study, Wong et al. [29] developed a novel DL-based CT image denoising method where the model was trained on noncontrast head CT in patients with acute ischemic stroke. Each CT scan was unique regarding imaging protocol, scanner vendor and model, radiation dose, etc. A limitation of their study is that they only included three cases in the assessment of stroke visualization and that they did not quantify the direct visual comparisons. Furthermore, the visual comparisons were not performed by radiologists. Also, the denoising performance assessment was only performed using the top 30 images of the supratentorial region. The strengths with our study compared to that of Wong et al. are that we included quantitative image quality parameters on several anatomical levels and also evaluated qualitative parameters in a quantized and standardized manner. Furthermore, our qualitative image quality assessment was performed by end-users.

Our study confirms the findings by Oostveen et al., who detected lower image noise and improved gray-white matter differentiation of DLIR compared to IR-based image reconstruction algorithms for noncontrast head CT [25]. However, they included patients who underwent a head CT scan for various indications and did not incorporate intracranial pathology in their image quality assessment. Also, our sample size was nearly twice as large.

Our results are also in line with a review by Arndt et al. who evaluated DLIR phantom and body studies and concluded that DLIR algorithms improve image quality with the potential for radiation dose reduction [30]. Another review by Zhang et al. confirms that DLIR preserves image quality better at low doses compared to other image reconstruction techniques [20]. However, the nonlinear properties of DLIR algorithms can occasionally cause complex and unanticipated effects on image quality parameters. One study has reported that even though DLIR algorithms achieve less noise than FBP or IR, the spatial resolution can become degraded, especially with a decreasing radiation dose of the training images for the DLIR algorithm [31]. Furthermore, a phantom study by Solomon et al. showed superior low-contrast resolution of FBP compared to both ASiR-V and TrueFidelity and the low-contrast resolution decreased with increasing ASiR-V and DLIR strengths [32]. Another study by Jensen et al. reported blurring of small abdominal lesions that increases with higher DLIR strengths [33], rather analogous to what has been observed with higher IR strengths [14‐16]. We did not find any degradation of the image quality at higher DLIR strengths. We detected subtle, yet statistically significant differences in CT attenuation between the reconstructions (Table 2). A further analysis of the mean absolute HU differences as pairwise differences between all possible combinations of the four image reconstruction types (Supplementary Table 3) showed that the HU differences were slightly greater between ASiR-V and any DLIR strength level than between any possible combination of DLIR strength pairs, for all four brain regions. Additionally, a greater number of cases with statistically significant mean absolute HU differences were detected between pairs where one of the reconstruction types in one or both pairs was ASiR-V (37 cases), compared to if it was DLIR-L (16 cases), DLIR-M (15 cases), or DLIR-H (14 cases). These findings indicate that DLIR might cause minimal alterations in the attenuation values (in our case mean absolute HU differences of up to 1.03 HU). The clinical consequences of this finding are most likely insignificant.

There was a substantial variability in inter-reader agreement. This could partially be due to the implementation of a 5-grade scoring system. A scoring system with fewer scoring levels could have resulted in a higher inter-reader agreement, at the expense of less specific qualitative image scores. Another contributing factor to the variability in κ_w-values could be that the scores for artifacts and ICH conspicuity for the most experienced reader were relatively uniform between the different image reconstructions compared to the second most and the less experienced readers. This could also indicate that the superior image quality of DLIR did not influence the diagnostic confidence of the most experienced reader to the same degree as it did for the second most and the less experienced readers. Further studies are needed to evaluate if DLIR increases the diagnostic confidence for trauma head CT to a greater extent in less experienced compared to more experienced readers. If this is the case, it could lead to improved acute diagnostic accuracy for head CT for trauma patients at our trauma unit as the preliminary interpretation of the trauma CT usually is performed by our radiology residents (which is probably true for many trauma units during on call hours).

There are several limitations of our study. First, patients were evaluated retrospectively at a single level 1 trauma center. Second, the study did not directly assess diagnostic performance, which is necessary to fully appreciate the potential clinical benefits of DLIR. Third, images were obtained by a standard dose CT protocol. Hence, the radiation dose reduction potential of DLIR could not be directly determined. Fourth, assessment of normal anatomical structures was also performed in cases with ICH. Even though the evaluation in these cases was performed on the unaffected contralateral hemisphere, confounding effects on image evaluation due to the pathology cannot be fully excluded. Fifth, even though a spontaneous intracerebral hemorrhage is not a trauma-related finding per se, we decided to include this ICH type in the study as it is found in a small fraction of the trauma patients that present to our trauma unit and therefore mirrors the clinical scenario. Sixth, the CT scans were obtained with our standard trauma CT protocol with a slice thickness of 0.625 mm with 0.3125 mm overlap. However, the applied version of TrueFidelity on the raw data did not offer overlapping reconstructions. We are confident that this difference did not significantly affect the image quality comparison between ASiR-V and DLIR. If anything, the overlapping slices could potentially improve the image quality of ASiR-V by increasing the detection of small lesions [34]. We did not alter our standard trauma CT protocol by removing the slice overlap because we wanted to reflect the clinical scenario at our department as far as possible as this project also served the purpose as quality control for a novel trauma CT protocol. Seventh, we only evaluated how DLIR performs compared to ASiR-V, an IR technique based on hybrid/adapted-IR. We did not evaluate how DLIR performs compared to full/model-based IR techniques (MBIR). A phantom study by Higaki et al. has demonstrated that the image quality achieved by DLIR at low radiation doses outperforms that achieved by MBIR, but that MBIR outperforms DLIR at high radiation doses [35]. In a review, Nakamura et al. stated that DLIR may enhance the detection of low-contrast lesions compared to MBIR. They also indicated that MBIR enhances the detection of high-contrast lesions at high radiation doses compared to DLIR [36]. Further studies are needed to evaluate the clinical implications of these findings. Eighth, the calculated SNRs in our study do not represent true SNRs. True signal is the output signal of the detector, which, for energy-integrating detectors, is proportional to the integrated energy levels of all received photons [37]. The HU in turn is the linear transformation of the measured linear attenuation coefficient [38]. It would have been sufficient only to present the image noise. However, SNR, as defined in this article, is a frequently encountered quantitative CT image quality parameter in the literature, probably recognizable by many readers and easily related to.

Our findings are in line with the current abovementioned literature and extends the superior image quality of DLIR for head CT to the trauma setting as well. Further studies are needed to evaluate the performance of DLIR for other anatomical locations and spectrum of traumatic injuries as well as to determine if DLIR increases the diagnostic confidence to a higher degree for less experienced than for more experienced readers. A next step after that would be to determine the degree of radiation dose reduction that DLIR allows while maintaining diagnostic image quality in trauma patients.

In conclusion, the image quality of trauma head CT series reconstructed with DLIR outperformed CT series reconstructed with ASiR-V. In particular, DLIR-M and DLIR-H demonstrated significantly improved image quality and a lower proportion of non-diagnostic images. The qualitative image quality improvement provided by DLIR was more evident for the second most and the less experienced readers compared to the most experienced reader.