In oncological settings, the detection of lytic or sclerotic bone lesions is an integral part of computed tomography (CT) analysis. In 80% of instances, these metastatic bone lesions result from prostate, breast, and lung cancer [1], [2], [3], [4], [5]. The detection of bone metastases is crucial, because it affects staging and prognosis and can have important therapeutic implications [6]. However, the radiation dose delivered to patients to detect these lesions for lumbar spine CT examinations is high. Many tools have been developed in recent years to reduce the dose delivered to patients such as tube current modulation (TCM) systems and iterative reconstruction (IR) algorithms [7]. Using these two tools, acquisitions of lumbar spine CT images are performed in our institution on a routine basis with low-dose for all patients [8]. An ultra-low dose CT protocol is also used for patients presenting to the emergency department with post-trauma fractures [9].

Recently, new CT image reconstruction algorithms based on deep learning (DLR) have been developed by three major manufacturers [10]. These DLR algorithms feature a deep neural network or convolutional neural network trained to differentiate signal from noise from patient and/or phantom images reconstructed with filtered back-projection or model-based iterative reconstruction algorithms. They reduce image noise and preserve image texture, which is modified (image smoothing) with IR algorithms. They also have a higher dose reduction potential than IR algorithms. Several studies have shown that the two first commercially available DLR algorithms achieved dose reduction while maintaining diagnostic image quality [11], [12], [13], [14], [15], [16], [17], [18], [19], [20].

The DLR Precise Image algorithm (Philips Healthcare) is used in clinical routine for CT protocols of the thorax, abdomen and pelvis. The quality of parenchymal images with a "Lung" reconstruction kernel and mediastinal and abdominal-pelvic images with a "soft tissue" kernel has been previously validated in phantom and patient studies [21,22]. However, the impact of this new algorithm on lumbar spine images with a "Bone" kernel has not been evaluated.

The purpose of our study was to assess the impact of this new DLR algorithm on image quality compared with a hybrid IR algorithm in lumbar spine CT. A task-based image quality assessment and a subjective image quality assessment were conducted to achieve this goal.