Evaluation of deep learning image reconstruction for brain- and abdominal CT - A visual grading characteristics study

Abstract

Computed tomography (CT) is a common medical imaging method today. There are different ways of reconstructing the images from a CT examination. The standard method is the filtered back projection (FBP). However, a disadvantage of FBP is the need of a relatively high exposure in order to reduce the noise in the reconstructed images, which leads to high radiation doses to the patients. Therefore, iterative reconstruction (IR) methods have been developed which use statistical CT imaging models to estimate and reconstruct an image. This method reduces the radiation dose to the patient but also changes the image texture. Deep learning image reconstruction (DLIR) is a new reconstruction method based on deep neural networks (DNN). It has been shown that images reconstructed using DLIR have similar image texture as images reconstructed using FBP, while the image noise is reduced. At Sahlgrenska University Hospital, there are three CT-systems (GE Revolution Apex™ CT, GE Healthcare, Milwaukee, USA) equipped with True Fidelity which is a DLIR software. The present study aims at evaluating the image quality for DLIR with different post processing filters compared to ASIR-V for abdominal and brain examinations. Images from 20 abdominal examinations and 20 brain examinations were collected retrospectively and reconstructed. The reconstructions made for the abdominal examinations were; Stnd ASIR-V 40% 3 mm, Stnd DLIR-M + E1 3 mm, Stnd DLIR-M + E1 0.625 mm, Stnd DLIR-H + E1 3 mm and Stnd DLIR-H + E1 0.625 mm. The reconstructions made for the brain examination were; Soft ASIR-V 50% EC1 3 mm, Stnd DLIR-M EC1 3 mm, Stnd DLIR-M EC2 3 mm, Stnd DLIR-H EC1 3 mm and Stnd DLIR-H EC2 3 mm. The reconstructed images were evaluated in a visual grading characteristics (VGC) study where the images were rated according to five image quality criteria. Three abdominal radiologists and three neuro radiologists participated in the study. Three phantom studies were also performed. A Catphan phantom was used to evaluate if the spatial resolution, visibility of low contrast objects and linearity of HU were affected by the different reconstructions. An anthropomorphic phantom was used to show the effect on image quality when lowering the radiation dose. Lastly, a water phantom was used to compare the noise properties between the different reconstructions. The results from the VGC study of the abdominal examinations showed that the images reconstructed using DLIR were rated significantly higher than the images reconstructed using ASIR-V 40%. The abdominal radiologists rated the DLIR-H E1 3 mm reconstruction the highest. However, for the brain examination, most of the images reconstructed using DLIR were rated similar to the images reconstructed using ASIR-V 50%. The phantom studies showed that there were no significant differences for the spatial resolution, visibility of low contrast objects or linearity of HU between the different reconstructions. However, it was shown that the noise magnitude was lower in the images reconstructed using DLIR compared to the images reconstructed using ASIR-V. The overall result showed that there might be a possibility to reduce the radiation dose for abdominal examinations when reconstructing the images using DLIR instead of ASIR-V.