Network based systems

Virtual thin slice technique based on a three-dimensional conditional generative adversarial network for morphological assessment of the spine

This retrospective study was approved by the Osaka University Clinical Research Review Board, and the informed consent requirement was waived by the University Clinical Research Review Board. from Osaka. All methods were performed in accordance with current guidelines and regulations. Patients who underwent a CT scan for the evaluation of aortic or cardiac disease were eligible for inclusion in this study because we obtained a single breath-hold CT scan of the supraclavicular area at the symphysis pubis in these patients , while separate scans were obtained for the chest and abdominopelvic regions in other patients. Enrolled were 73 consecutive patients who underwent CT between January and February 2019 or between December 2020 and January 2021 (50 men and 23 women; age range, 25-91 years; mean age, 72.9 years). The clinical indications for CT in these patients are listed in Table 1.

Table 1 Clinical indications for CT in included patients.

CT examination

Computed tomography was performed using a 160 or 320 slice scanner (Aquilion Precision, Canon Medical Systems, Otawara, Japan, n=34, or Aquilion ONE GENESIS Edition, Canon Medical Systems, n=39). A pre-contrast CT scan was performed in all patients from the supraclavicular region to the symphysis pubis during a single apnea. The tube current was individually adjusted using a self-exposure control technique with a standard deviation setting of 15. The remaining scan parameters were: tube voltage, 120 kVp; rotation time, 0.5 sec; helical pitch, 0.83. Although post-contrast scans were also acquired in 31 patients, only pre-contrast images were used in this study.

From the raw data of each patient, two sets of axial images were reconstructed, with a slice/interval thickness of 4/4 and 1/1 mm. A hybrid iterative reconstruction algorithm (AIDR 3D, Canon Medical Systems) with a low force setting was applied. The remaining reconstruction parameters were: core, FC03; reconstruction field of view, 350 mm (pixel size, 0.68 × 0.68 mm).

Virtual thin slice technique

VTS is a conditional GAN ​​based algorithm. Thick slice images with 3-10 mm slice thickness/intervals were randomly simulated from real thin slice images by subsampling with Gaussian smoothing. A pair of original thin-slice images and simulated thick-slice images were used to train the VTS generator in the GAN framework (Fig. 1). The generator is an encoder-decoder type architecture with connection hops inspired by U-Net to reconstruct high resolution images. The role of the discriminator is to allow the generator to produce virtual images in thin slices which are difficult to distinguish from the real ones. Both the generator and the discriminator are composed of 3D convolutional neural networks. Conditioning labels (eg, slice interval) associated with input thick images are fed into the discriminator to improve super resolution accuracies. During generator training, the L1 loss was calculated in addition to the adversarial loss, in order to minimize the intensity difference at the pixel level between the original (ground truth) and the generated thin-slice images, since those these should be as close as possible. VTS software is a feature of PACS Viewer (SYNAPSE SAI Viewer Version 1.0, FUJIFILM, Tokyo, Japan), which has received regulatory approval in Japan. The training CT data for this software contained CT images of different body parts (head, chest, abdomen and legs) obtained with scanners from different manufacturers. Thus, the software can be applied to any part of the body. The generated VTS images were isotropic with a voxel size of 1 × 1 × 1 mm. Details of the VTS technique were presented at a previous conference, and the manuscript is available for reference on the preprint server14. VTS software was applied to each patient’s 4 mm thick dataset to generate 1 mm thick VTS images.

Figure 1

Contradictory training framework for thick-thin slice translation of CT images.

Qualitative analysis

Two radiologists familiar with abdominal radiology (9 and 6 years of experience) independently reviewed sagittal images reformatted from 4 mm thick images and VTS images and assessed the visibility of intervertebral spaces in each of the four regions: cervical, upper thoracic, lower thoracic and lumbar spine. They reviewed these images on a commercially available workstation (SYNAPSE VINCENT version 5.3.001, FUJIFILM) and assigned a score using the following 4-point scale: 4, all intervertebral spaces visible ; 3, most of the intervertebral spaces are visible but some are not clear; 2, most intervertebral spaces are unclear; 1, no intervertebral space is visible. The radiologists were informed that the images to be evaluated were either 4 mm thick images or VTS images, but did not know the identity of the patients, their medical history and the reconstruction protocol used.

Quantitative analysis

Two radiologists familiar with abdominal radiology (16 and 9 years of experience), different from the radiologists who performed the qualitative assessment, independently measured the height of the first thoracic vertebra (Th1) and the first lumbar vertebra (L1) on Sagittal reformatted images made from each of the 4 mm thick, true 1 mm thick, and VTS datasets. The height was measured at the anterior edge of each of these vertebrae. The absolute values ​​of the difference between the heights measured on the images of 4 mm thickness and real ones of 1 mm thickness (D1) were calculated, as well as the absolute values ​​of the difference between the heights measured on VTS and the real images of 1 mm thickness (D2). The absolute errors in percentage between the heights measured on the images of 4 mm thickness and the real images of 1 mm thickness (% Error1) was also calculated by dividing D1 by the height measured on the real images of 1 mm thickness, as well as the absolute errors in percentage between the heights measured on the VTS and the real images of 1 mm thickness (% Error2). The measurements were carried out using a workstation (SYNAPSE VINCENT version 5.3.001).

Diagnostic performance in the detection of compression fractures

The same two radiologists who performed the qualitative assessment also independently assessed the possible presence of a compression fracture using the reformatted sagittal images constructed from each of the 4 mm thick images and the images VTS. They ranked the probability of compression fracture in all vertebrae using the following 4-point confidence score scale: 1, probably no fracture present; 2, indefinite presence of fracture; 3, fracture probably present; and 4, fracture clearly present. Prior to assessment, they were informed that a confidence level of 3 or 4 would be considered a positive result for the calculation of sensitivity and positive predictive value (PPV). The compression fracture criteria used in this study were: 1, ratio of anterior vertebral height (AH) to posterior height (PH) 20% compared to those above and below15. The reference standard was determined by two other radiologists (16 and 9 years of experience) who evaluated the presence or absence of a compression fracture on sagittal images reformatted from the true images of 1 mm thickness, in consensus.

statistical analyzes

Visual scores for intervertebral space visibility were compared using the Wilcoxon signed rank test. The absolute values ​​of the difference in the measured vertebral heights (D1 and D2) were compared using you-test. The absolute errors in percentage of the measured vertebral heights (%Error1 and %Error2) were also compared using you-test. Interobserver agreement for each of the D1 and D2 was assessed by the intraclass correlation coefficient (ICC). To analyze the diagnostic performance of compression fracture detection, an analysis of jackknife free-response receiver operating characteristics (JAFROC) was performed using JAFROC software (JAFROC Version 4.2.1, www.devchakraborty.com). This software calculates the Factor of Merit (FOM), which is defined as the probability of a lesion being scored higher than the highest-scoring non-lesion on a normal image16. In the present study, JAFROC1 was used rather than JAFROC or JAFROC2 due to its high statistical power for human observers.17. For all tests, a P a value less than 0.05 was considered significant.


Source link