Comparison of image quality between Deep learning image reconstruction and Iterative reconstruction technique for CT Brain- a pilot study

Introduction

Computed tomography (CT) is the primary imaging modality used to evaluate patients suspected to have central nervous system disorders. The ability to visualize brain regions quickly and thoroughly is one of their main advantages. This helps in the timely assessment of problems, including stroke, trauma, and intracranial lesions. Non-contrast CT (NCCT) brain scans are widely used in a variety of therapeutic contexts because of their accessibility, speed, and efficacy, which are vital for the early diagnosis and treatment of neurological diseases.^1–3

One notable development in CT reconstruction technology is the use of iterative reconstruction (IR) techniques. iDose⁴ is a 4^th generation IR method released by Philips Healthcare that offers improved image quality (IQ) with a reduced radiation dose. IQ and diagnostic precision are improved through IR, which uses sophisticated mathematical techniques to optimize and refine image data. A significant reduction in image noise (IN) allows for clearer visualization of anatomical structures, especially in regions of low contrast, which is the main advantage of the IR technique. IR allows the acquisition of high-quality images with a decreased radiation dose (RD) for patients. For vulnerable groups such as children or those who are radiation-sensitive, this is especially important.^4,5 The IR technique results in a plastic or blotchy appearance at higher reconstruction levels.^6–8

Deep learning image reconstruction (DLIR) algorithms represent a transformative leap in image reconstruction and dose reduction in CT. DLIR in Philips Healthcare is called a Precise Image (PI). Philips PI is the latest and most reliable way to reconstruct high-quality CT images using artificial intelligence (AI) techniques. A trained deep learning neural network was used in the PI reconstruction process. With the fastest reconstruction speed in the market, PI preserves the traditional view of FBP photos.⁹ By harnessing the capabilities of artificial intelligence (AI), particularly convolutional neural networks (CNNs), these algorithms revolutionize image reconstruction by learning intricate patterns from raw CT data. These algorithms learn complex relationships between sparse or noisy input projections and corresponding standard-dose reference images, enabling the generation of clinically acceptable reconstructions even when using a reduced radiation dose. By leveraging the inherent information within the data and learning intricate patterns, DLIR contributes to the advancement of dose reduction strategies in CT. The DLIR technique yields an image texture reminiscent of FBP, even at low-dose strengths.^10–12 There are limited studies on the usefulness of DLIR on IQ in the NCCT brain. Hence, the aim of this study was to compare the IQ between the new Precise Image (DLIR) and IR (iDose⁴) techniques for the NCCT Brain.

Methods

This is a prospective study. The Institutional Ethical Committee (IEC 400/2022) was obtained from Kasturba Medical College and Hospital, Manipal, India, on 1^st July 1, 2023, followed by the Clinical Trial Registry - India (CTRI) registration (CTRI/2023/07/055310) on 18^th July 2023. Written informed consent was obtained from all the participants for publication and participation in the study.

Eligibility criteria: Thirty patients referred for the NCCT brain were included. Patients referred to the NCCT brain for various clinical indications such as trauma, seizures, stroke, headache, vomiting, fever, chills, and breast carcinoma were included. Patients who were uncooperative and who underwent CT scans with motion artifacts were excluded. The patients included in the study had neuropathological findings on CT, including hemorrhage (n=10), infarct (n=10), tumor (n = 5), VP shunt (n = 1), arachnoid cyst (n = 1), metastases with edema (n=1), cerebral atrophy (n=1), and encephalomalacia (n=1).

CT Image acquisition: This study was performed at the Department of Radiodiagnosis, Kasturba Medical College and Hospital. Patients referred for NCCT brain examinations underwent 128 slice CT (Incisive, Philips Healthcare). The technical parameters for the NCCT brain acquisition are listed in Table 1. The images were reconstructed using iDose⁴ (level 3)⁴ and the DLIR reconstruction⁹ level standard.

“Qualitative Image quality analysis”: Two radiologists [reader 1 (R1) and reader 2 (R2)] with over 15 years of experience in neuroradiology imaging evaluated the CT images. Both the readers were blinded to the reconstruction level. The readers assessed the “Overall image quality” (OQ), “Image noise” (IN), “Artifacts” using 5-point Likert scale (Figure 1).

Figure 1. Showing the 5-point Likert scale for qualitative image quality analysis.

“Quantitative Image quality analysis”: CT attenuation (HU) and image noise (IN) of gray matter (GM) and white matter (WM) at the level of the basal ganglia (BG) and centrum-semiovale (CSO) regions were measured. To calculate attenuation (HU) at the GM and WM of the BG, an ROI of 0.1-0.2 cm² was placed in the thalamus and posterior limb of the internal capsule (PIC). For calculating attenuation at GM and WM of CSO, the ROI of 0.1-0.2 cm² was placed in the region of frontal WM and adjacent cortical GM (Figure 2a-b).

Figure 2. Axial CT images (DLIR-standard) of 61-year old male at the level of centrum semiovale (a) and basal ganglia (b) showing region of interest (ROI) in gray matter and white matter. Axial CT image at the posterior cranial fossa (c) with ROI drawn in the pons region between the petrous bone.

The posterior fossa index (PFI) was calculated as the image noise (Standard deviation-SD) of the HU values in the pons. To calculate the PFI, an ROI of 0.2-0.3 cm² was placed in the pons region of the posterior cranial fossa (Figure 2c).

The signal-to-noise ratio (SNR) at the BG and CSO levels was calculated as mean CT attenuation (HU)/standard deviation (SD) (SD: Image noise).

The contrast-to-noise ratio (CNR) at the BG and CSO levels was calculated using the following formula:

Radiation dose metrics such as “CTDI_volume (CTDI_vol), dose length product (DLP), and size-specific dose estimate (SSDE)” were recorded from the display of the console monitor.

Results

A total of thirty patients with 22 males and 8 females with mean age of 55.46±15.38 years referred for NCCT brain were included (Table 2). The mean CTDI_vol, DLP, and Size specific dose estimate (SSDE) were 46.36±0.20 mGy, 1157.5±64.23 mGy.cm and 44.10 mGy respectively.

Qualitative IQ analysis

Qualitative IQ analysis showed an increase in scores for OQ, IN, and artifacts with DLIR-standard compared to iDose⁴ for both readers (Table 3) (Figure 3a-b).

Table 3. Showing qualitative image quality analysis between iDose⁴ and DLIR-standard.

IQ	iDose4			DLIR-standard			p-value
	R1	R2	k	R1	R2	k
OQ	3.2±0.4	3.2±0.4	0.902	4.2±0.49	4.13±0.43	0.821	<0.05
IN	3.03±0.41	3.03±0.41	1.00	4.16±0.46	4.23±0.53	0.837	<0.05
Artifacts	3.3±0.59	3.33±0.47	0.80	4.23±0.67	4.26±0.63	0.943	<0.05

Figure 3. Axial CT images showing the improved image quality with (a) iDose⁴ compared to (b) DLIR-standard.

The OQ showed a significant difference (<0.05) between the iDose⁴ (3.2±0.4; R1) and DLIR-standard (4.2±0.49; R1). IN showed a significant difference (<0.05) between iDose⁴ (3.03±0.41; R1) and DLIR-standard (4.16±0.46; R1). Artifacts showed significant differences (<0.05) between iDose⁴ (3.3±0.59 R1) and DLIR-standard (4.23±0.67 R1) (Table 3).

Discussion

In the present study, we compared the qualitative and quantitative IQ between the DLIR-standard (Precise Image) and IR (iDose⁴) techniques for the NCCT brain. Our study noticed that both qualitative and quantitative IQ improved significantly with the DLIR-standard compared with the iDose⁴. The new DLIR technique, Precise Image, outperformed the IR technique (iDose⁴). Our study found that the DLIR standard showed a significant reduction in IN and an increase in SNR and CNR at BG and CSO levels. The DLIR-standard showed higher subjective IQ scores with excellent agreement between readers compared to the iDose⁴. The lower IN, higher SNR, and CNR might allow for lowering the radiation dose with the DLIR-standard compared to iDose⁴ for the NCCT brain.

Studies by Kim et al.¹³ and Alagic et al.¹⁴ showed 24-52% and 3.5-43% reduction in IN for NCCT brains with DLIR (True Fidelity; GE) reconstruction levels of low, medium, and high compared with ASIR-V (Adaptive statistical iterative reconstruction-Veo) at BG and CSO levels. DLIR-standard in the present showed a 41.8-47.6% reduction in IN at BG and CSO levels, which is similar to the results of Kim et al.¹³ and Alagic et al.¹⁴ Another Two studies by Oostveen et al.¹⁵ and Cozzi et al.¹⁶ reported a 9.6% and 13% reduction in IN for NCCT brains with DLIR (AiCE) compared with “hybrid-iterative reconstruction (Hybrid-IR)” and “model-based iterative reconstruction” (MBIR), “Adaptive iterative dose reduction” (AIDR-3D), which is slightly less IN reduction compared to our study. The slight variation in the reduction of IN across CT vendors might suggest the need for further research comparing different reconstruction algorithms.

For the NCCT brain, diagnostic evaluation of the posterior fossa in emergency situations to identify hemorrhagic and ischemic events is important. However, the posterior cranial fossa often experiences beam hardening, streak, and partial volume artifacts due to the presence of bony structures surrounding the cerebellum, pons, and medulla, which leads to diagnostic challenges in identifying hemorrhage and infarct in this region. The artifact index could indicate the extent of CT number fluctuations resulting from artifacts, along with intrinsic image noise linked to factors related to both the scanner and the patient.^17,18 Our study noticed a 27.08% reduction in IN and artifact index in the pons region of the posterior fossa with DLIR-standard compared to iDose⁴ which was similar to the artifact index reported by Kim et al.¹³ (17-38%) and Alagic et al.¹⁴ (6.8-32.8%). However, Cozzi et al.¹⁶ reported a higher artifact index (median 8.4; interquartile range 7.3-9.2) with DLIR (AiCE) than with AIDR-3D (median 7.5; interquartile range 6.9-8.3) in thin sections, which is contrary to the study reported by Oostveen et al.¹⁵ with the same DLIR technique. The reason for this could be the difference in the placement of the ROI and slice thickness used between the two studies.

Our study observed higher SNR (65-82%) at BG and CSO levels with DLIR-standard, which suggests better gray and white matter differentiation compared to iDose⁴. The findings of our study were similar to the results of Alagic et al.¹⁴ (2-89%) with DLIR-low, medium, and high levels, and Pula et al.¹⁹ (46-59%) with DLIR-High compared to IR techniques. A study by Oostveen et al.¹⁵ reported a slightly lower reduction in SNR (5-26%) compared to our study because of the difference in the formula used for calculating the SNR. Our study observed an increase in CNR (68-78.8%) with DLIR-standard at BG and CSO levels, which suggested better gray and white matter differentiation compared to iDose⁴. The results of our study were similar to the findings reported by Alagic et al.¹⁴ (2.4-53%) and Cozzi et al.¹⁶ (28-39%). However, Kim et al.¹³ reported an increase in CNR of 99% with DLIR-high DLIR.

Our study found no significant difference (p>0.05) in CT attenuation (HU) between DLIR-standard and iDose⁴ at the BG and CSO levels. The findings of our study are similar to the results of Kim et al.¹³ which showed no significant difference in CT attenuation of GM between DLIR levels and IR technique. However, a study by Alagic et al.¹⁴ reported significant differences in CT attenuation between DLIR levels and IR technique at the PLIC WM and adjacent cortical GM, which might suggest that DLIR could lead to minor changes in attenuation values; however, this finding is unlikely to have significant clinical consequences.

The DLIR-standard showed higher qualitative scores for OQ, IN, and artifacts compared to iDose⁴ which is similar to the findings reported by Oostveen et al.¹⁵ and Pula et al.¹⁹ Studies by Kim et al.¹³ and Alagic et al.¹⁴ reported an increase in qualitative scores with an increase in the strengths of DLIR from low to high compared to IR techniques.

Our study has a few limitations. First, the study involved a small sample size, and it is necessary to conduct further research with a larger patient cohort to confirm our study findings. Second, the study did not directly evaluate the diagnostic efficacy, which is a crucial step in understanding the complete clinical advantages of DLIR. Third, CT scanning was performed using a standard dose protocol, making it challenging to directly ascertain the potential dose-reduction benefits of DLIR.

Conclusion

The New DLIR Precise Image (DLIR) technique offers improved image quality with reduced image noise and higher SNR and CNR than iDose⁴. The DLIR standard also showed higher qualitative image quality scores than the iDose⁴. The reduction of posterior fossa artifacts with the DLIR standard for the NCCT brain improves the diagnostic accuracy of identifying hemorrhages/infarcts in emergency cases. Our current study may provide implications for performing low-dose scans with reduced radiation doses using DLIR in the NCCT brain.

Ethics and consent

The Institutional Ethical Committee (IEC 400/2022) was obtained from Kasturba Medical College and Hospital, Manipal, India, on 1^st July 1, 2023.

Written informed consent was obtained from all the participants for publication and participation in the study.