Deep learning–based reconstruction may improve non-contrast cerebral CT imaging compared to other current reconstruction algorithms

Cerebral NCCT scans in a consecutive cohort of 50 patients that underwent NCCT between 14 June 2019 and 7 September 2019 for various clinical indications were collected. Two scans were incorrectly labelled and appeared to be contrast scans and were excluded. Of the remaining 48 patients, 28 were male, the age range was 19–86 years with a median age of 66 years. During clinical interpretation, intracranial haemorrhage was found in 19 patients, 13 patients had signs of infarct, and a tumour was noted in 4 patients. Seven patients had a ventricular shunt in place, four patients had a coiled aneurysm, two had an intracranial pressure monitor, and one had a stereotactic frame.

All acquisitions were performed on a 320 row-detector CT scanner (Aquilion One PRISM edition, Canon Medical Systems Corporation). Patients were scanned using different scan modes and techniques, as listed in Table 1.

All acquisitions were reconstructed using three reconstruction techniques: Hybrid-IR (AIDR 3D Enhanced, brain-kernel FC26), MBIR (FIRST Brain Standard), and DLR (AiCE Brain LCD). All reconstructions had a slice thickness of 0.5 mm, and the same field of view.

The image quality of the 48 reconstructed NCCTs was evaluated by four experienced radiologists and one final-year resident specialising in neuroradiology. Noise magnitude, sharpness, natural appearance, artefacts, grey-white matter differentiation and overall perceived image quality were scored using a 5-point Likert scale, as detailed in Table 2. The assessments were performed on a workstation with calibrated diagnostic screens (Barco MDNC-3321) in a radiology reading room with dimmed lighting. The observers were blinded to the reconstruction technique and images were presented in a random order. The order of the cases and reconstruction techniques differed for each observer.

Three ROIs were placed in each of the reconstructed volumes by an imaging scientist with 22 years of experience in x-ray imaging (L.O.), supervised by a neuro-radiologist (E.S.), encompassing different structures: cerebro-spinal fluid (CSF) within the lateral ventricle, centrum semiovale (white matter) and putamen (grey matter). The oval-shaped ROIs contained between 311 and 475 pixels. The standard deviation in the CSF (SD_CSF) was determined as a measure of noise magnitude, and the signal-difference-to-noise ratio (SDNR) of the grey and white matter was calculated using (HU_putamen – HU_{centrum semiovale})/SD_CSF, where HU is the mean of the Hounsfield Units in the voxels in the respective ROIs. Since the posterior fossa has more bony structures and therefore might be more prone to artefacts, the SD in the fourth ventricle was also measured, to determine if different values would be found compared to the measurement in the lateral ventricle.

In order to obtain a measure of the reconstruction duration, the total reconstruction time for each reconstruction method was measured manually for the first 15 acquisitions. Since the reconstructions were performed on the CT system, the reported reconstruction times are representative of those that can be expected in clinical practice.

Subgroup analysis was performed evaluating image quality of NCCT scanned in sequential and helical modes. The same method as described above was used, but a factor containing the acquisition type was added to the GLM. In order to see if artefacts appear differently across the three reconstruction methods, another subgroup analysis was performed using only scans of patients having foreign bodies.

Figure 2 shows stacked bar graphs for the observer ratings of each image quality characteristic across all cases and observers. In MBIR, a maximum of 20% of the ratings was scored as 4 or 5, while for DLR at least 50% achieved these ratings. The differences in ratings for DLR and Hybrid-IR are less distinct and vary across image quality characteristics. The estimated marginal means for each reconstruction method resulting from the GLM are shown in Table 3. For all image quality characteristics, DLR scores were significantly better compared to MBIR, while compared to Hybrid-IR perceived noise and grey-white differentiation were preferred with DLR. An example slice reconstructed with the three different reconstruction algorithms is provided in Fig. 3. The image noise is notably lower for DLR as compared to Hybrid-IR and MBIR, resulting in better differentiation between grey and white matter. Table 4 shows the results found when analysing the scans of only the patients having a foreign body. The scores did not show any significant differences in the observer preferences than those found for the entire case set.

The ratings of the different observers are shown in Fig. S1 (supplementary online material) using stack bar graphs per observer and per image quality characteristic. Dichotomising the scores per image quality characteristic between “DLR performs equal or better” and “DLR performs worse” when comparing DLR to the other reconstruction techniques results in a maximum 10% deviation from the mean for each observer compared to the overall mean across observers, except for the mean score of one observer when assessing the grey-white matter differentiation (15%), for one observer evaluating sharpness (− 31%), and two observers when evaluating natural appearance (36% and 31%). These four larger differences were found in the comparison between DLR and Hybrid-IR. Table 5 shows that all image quality characteristics were scored higher for NCCTs scanned in sequential mode in comparison to helical mode. Using this model, which includes acquisition technique, the reconstruction method remains a significant factor (p < 0.001), while the interaction between reconstruction method and acquisition technique was non-significant (p = 0.21).

The positions of the ROIs in the CSF within the lateral ventricle, centrum semiovale and putamen are shown in Fig. 4. The measured noise magnitude (SD_csf) of DLR was clearly lower than the noise magnitude for Hybrid-IR and MBIR, as can be seen in Table 3. DLR also resulted in the highest SDNR between grey and white matter. These improvements in image quality are obtained with a modest increase in reconstruction time compared to Hybrid-IR (+ 17 s), but with substantial reduction in reconstruction time compared to MBIR (− 132 s). The SD_csf and SDNR were significantly better for sequential scans than those of the helical scans, as shown in Table 5.