17th International Meeting on Fully 3D Image Reconstruction in Radiology and Nuclear Medicine 16 - 21 July 2023, Stony Brook, NY, USA Dirk Schäfer1, Fredrik Ståhl2,3, Artur Omar4,5, and Gavin Poludniowski4,6 1Medical Image Acquisition, Philips Research, Hamburg, Germany 2 Department of Neuroradiology, Karolinska University Hospital, Stockholm, Sweden 3Department of Clinical Neuroscience, Karolinska Institute, Stockholm, Sweden 4Medical Radiation Physics and Nuclear Medicine, Karolinska University Hospital, Stockholm, Sweden 5Department of Oncology-Pathology, Karolinska Institute, Stockholm, Sweden 6Department of Clinical Science, Intervention and Technology, Karolinska Institute, Huddinge, Sweden Abstract Dual-energy CBCT C-arm imaging is a new imaging modality, which may provide a variety of different spectral results in the interventional suite. We analyze model observers, i.e. the detectability index, of different infarct models, to get insight into which spectral reconstruction may be best suited for identifying infarcts prior to an intervention treating acute stroke patients. MTF and NPS measurements have been performed on a clinical dual-layer (DL) CBCT prototype system and used for model observer analysis. Infarct models with reconstruction dependent contrast amplitude are investigated using a variety of standard and spectral reconstructions. Clinical examples of subjects with small and large infarcts imaged on the same DL-CBCT system are compared visually and be means of CNR analysis of the infarcts to the results of the model observers. The model observer representing most closely a human reader prefers VMI images at higher keV values, and the same trend is obtained with simple CNR analysis. This result is partially substantiated by the CNR analysis of the clinical patient data. The prototype DL-CBCT system scans were conducted with 120 kV tube voltage, 2.5 mm aluminum-equivalent tube current . Both, standard and spectral images were generated from the DL-CBCT scans. The reconstruction of the front-layer (front) corresponds to the standard reconstruction of single layer C-arm system. The addition of front- and back-layer data yields the combined reconstruction (comb). The following spectral results were derived from materialdecomposed Compton scatter and photoelectric base functions [3]: Virtual monoenergetic image (VMI) and virtual non-contrast VNC images. Noise suppression in spectral images was realized by exploiting the anticorrelated noise behavior after material decomposition [4]. The denoising strength was selected by a clinical reader pilot study to yield best visualization of brain tissues for assessment of ischemic changes in stroke patients [5]. The modulation transfer function (MTF) was calculated from six scans of the head module of the multi-energy CT phantom model 662 (CIRS, Norfolk, VA, USA) using an established methodology [6], utilizing different inserts equivalent to various tissues, blood and iodine concentrations. Examples of the radially averaged MTFs for iodine blood inserts are shown in Fig. 1. Since they do not deviate appreciably, the low-contrast spatial resolution was assumed constant for a given reconstruction and the MTF curves for the 2 mg/ml I in blood was used subsequently. 1 Introduction Spectral dual-layer C-arm CBCT is a new imaging modality and may be used during stroke treatment workflow to improve visualization of infarct regions in the beginning of a procedure and support blood-iodine separation for hemorrhage identification at the end of the procedure. Presenting the most relevant information to the interventionalist is a challenging task. We aim to get insight into which spectral reconstruction is best suited to support a neuroradiologist in the task of identifying infarcts in CBCT reconstructions in acute stroke treatment. Reader studies based on dual-layer, dual-energy CT have shown optimal visualization of infarcts for Virtual Monoenergetic Image (VMI) reconstructions around 70 keV [1]. Availability of clinical data for dual-energy CBCT of acute stroke patients is limited, therefore we investigate the possibility to use model observers or measures such as CNR to support the choice of most suited spectral reconstructions for the infarct identification task. 2 Materials and Methods Measurements in this study were made with a noncommercial dual-layer CBCT prototype system [2]. The spectral DL-CBCT system is a a commercial interventional C-arm X-ray system with CBCT imaging capability (Allura Xper FD20/15, Philips Healthcare, The Netherlands), equipped with a dual-layer 20 inch (379.4 mm x 293.2 mm) detector prototype. The detector prototype consists of two detector layers stacked on top of each other. Figure 1. Radially averaged task-related modulation transfer functions (MTF) for Iodine blood inserts for VMI 75 keV. 353 17th International Meeting on Fully 3D Image Reconstruction in Radiology and Nuclear Medicine The 3D noise power spectrum (NPS) was calculated from six consecutive scans made of the homogenous water phantom for all reconstructions investigated in this study, using well-established techniques [7]. The 2D NPS was then calculated. A 2D NPS of the 75keV VMI reconstructions is shown in Fig. 2. 16 - 21 July 2023, Stony Brook, NY, USA perceived noise power spectrum. The quantity is and assumed to be real functions. The signal can be represented as [9]: , where is the 2D Fourier transform of the contrast profile C of the feature of interest, the MTF is the spatial resolution for the low-contrast task, and VTF is an assumed the visual transfer function (VTF) for a human Different modelled contrast profiles C have been investigated representing infarct regions of different size (5, 10 and 15 mm). The contrast amplitude was estimated based on different assumed percentages of water content mixed with gray matter brain tissue. Calculations of contrast were based on observed average CT numbers for water and gray matter substitutes for the different reconstructions. The resulting contrast and CNR values for standard and spectral reconstructions are shown for an infarct of 5 mm diameter with 5% water content in Fig 4. Figure 2. A 2D noise power spectrum corresponding to the trans-axial plane for VMI 75 keV. The model observers are evaluated on 2D trans-axial slices, corresponding to the clinical practice of reviewing reconstructions in acute stroke treatment. The trans-axial NPS is radially averaged, as shown in Fig. 3 for selected reconstructions, and used as on a resampled cartesian grid in the following. Figure 4. Modeled contrast, noise derived from NPS and resulting CNR used for model observers. In this study, the VTF due to Eckstein [10] is used with a peak sensitivity at approximately 4 cycles/degree: with , Figure 3. Radially averaged trans-axial 2D noise power spectrum for selected reconstructions where is the reconstruction field-of-view, R is the viewing distance (assumed to be 50 cm) and D is the display size of the FOV at a viewing station (assumed to be 25 cm). The (observed) noise power spectrum can be represented as, of the squared signal-to-noise ratio of a matched-filter model observer is given by [8]: . In this study, the matched filter will either be prewhitening or non-prewhitening [11]: where is the expected two-dimensional (2D) signal as perceived by the observer and is the 354 17th International Meeting on Fully 3D Image Reconstruction in Radiology and Nuclear Medicine or 16 - 21 July 2023, Stony Brook, NY, USA higher VMI. The results for a 5 mm infarct with water content of 10% and 15% simply scale up by a factor of 2 and 3, respectively (not shown). The results for 5% water content with lesion diameter of 10 mm and 15 mm are generally higher by a factor of approximately 2.5 and 4, respectively (not shown), but similar in trend with more pronounced maxima at VMI70 for pw and npw observer and similar flat npwe results slighlty increasing to higher energy levels of the VMI. , respectively. Note, that in the first case, the observer is able to decorrelate the perceived noise. This leads to cancelling and the numerator being equal to the square of the denominator. Three model observers will be explored: 1. PW: prewhitening matched filter. This corresponds to an ideal linear observer. 2. NPW: non-prewhitening filter. This corresponds to an observer unable to undo noise correlations. 3. NPWE: non-prewhitening filter with an eye filter. This corresponds to the NPW observer with the sensitivity of the human eye incorporated. Moreover, visibility of infarct regions in clinical examples of spectral DL-CBCT is assessed visually and by means of CNR analysis. The patient data has been acquired in a prospective single center clinical trial (NCT04571099), which enrolled consecutive patients, 50 years or older, with ischemic or hemorrhagic stroke on initial CT [5]. 3 Results The detectability index for the modelled infarct with 5 mm diameter and 5% water content is shown for the standard and spectral reconstructions in Fig 5. The prewhitening (pw) model observer values compared to the non-pw (npw) observer, but the relative trend is very similar with a maximum around VMI70. The results of the npw observer with eye filter (npwe) are sgnificantly reduced and the trend for different VMI is more flat with the maximum of shifted towards Figure 5. Results of different model observers for lesion size of 5 mm and water content of 5%. Clinical examples of a small and a large infarct are shown in Fig.6. The ROIs used to quantify the CNR of the infarct region with respect to gray matter brain tissue are indicated on the front layer reconstruction. The CNR values for both infarcts are plotted in Fig. 7. Fig. 6. Example of subject with a hyperacute, small infarct region (top row) and an acute large infarct region (bottom row). ROIs for measuring CNR are indicated in the front-layer reconstructions (left) and values reported in Fig. 7. All reconstructions are displayed with level / window of 25 / 70 HU and slice thickness of 5 mm . 355 17th International Meeting on Fully 3D Image Reconstruction in Radiology and Nuclear Medicine 16 - 21 July 2023, Stony Brook, NY, USA approximated as a signal known exactly/background known exactly scenario. 5 Conclusion This study presents the use of model observers to investigate different spectral reconstructions for infarct visualization in spectral DL-CBCT. The model observer representing most closely a human reader prefers VMI images at higher keV values, and the same trend is obtained with simple CNR analysis. This result is substantiated by the CNR analysis of a large infarct, while the clinical example of a small infarct showed less dependence on the energy level of the VMI reconstruction. These results have to be verified by clinical reader studies, but this framework may help to investigate and optimize system performance for different tasks with limited availability of clinical data. Fig. 7. CNR of the infarct/GM ROIs shown in Fig. 6 4 Discussion There is a general observation that VMI and VNC reconstructions provide better CNR and detectability index for the task of infarct visualization compared to front-layer and combined reconstructions (see Figs. 4, 5). The CNR values of the clinical example (Fig. 7) show a similar behavior for energy levels > 50 keV. This is partly caused by the additional denoising involved in the spectral reconstructions and proper separation of these effects is not trivial and beyond the scope of this paper. Interestingly, the optimal energy level of the VMI reconstruction for this task depends on the model observer. The non-prewhitening filter with an eye filter (npwe) is expected to match best to the performance of a human observer. The Npwe observer (Fig.5) and CNR (Fig.4) indicate higher mono-energy levels as optimal. Optimal CNR values for the clinical example with large infarct are also obtained for higher energy levels of the VMI, whereas the small infarct shows relatively flat dependence. The different CNR trends for the different clinical cases indicate that the composition of infarcts merits further attention as well as any other potential confounding factors. These investigations aim to identify the optimal spectral reconstruction to be presented for the task of infarct detection, which shall be cross checked in future work with reader studies on clinical data. The detectability index can be related to the area-under-the-curve (AUC) from receiver operating characteristics (ROC) studies or the percentage correct (PC) responses in multiple alternative forced choice (AFC) studies: , where is the standard cumulative normal distribution. The inclusion of so-called internal noise of the observer into the model may be necessary to quantitatively predict human performance [12]. Potential limitations of the approach include the assumption of quasi-linearity (permitting image analysis in the Fourier domain) and the assumption that the clinical task can be References [1] Ståhl F, Gontua V, Almqvista H, Mazyab MV, Falk Delgado A., Performance of dual layer dual energy CT virtual monoenergetic images to identify early ischemic changes in patients with anterior circulation large vessel occlusion, Journal of Neuroradiology, https://doi.org/10.1016/j.neurad.2020.12.002, 48, 75 81, 2021 [2] Ståhl F, Schäfer D, Omar A, van de Haar P, van Nijnatten F, Withagen P, Thran A, Hummel E, Menser B, Holmberg Å, Söderman M, Falk Delgado A, Poludniowski G. Performance characterization of a prototype dual-layer cone-beam computed tomography system. Med Phys 48(11):6740-6754. doi: 10.1002/mp.15240, 2021. [3] Shapira N, Yagil Y, Wainer N, Altman A. Spectral imaging technologies and apps and dual-layer detector. In: Taguchi IBK, Iniewski K , eds. Spectral, Photon Counting Computed Tomography: Technology and Applications. CRC Press; 3-16, 2020: [4] Brown KM, Zabic S, Shechter G, Impact of spectral separation in dual-energy CT with anti-correlated statistical reconstruction. Proceedings of the 13th fully three-dimensional image reconstruction in radiology and nuclear medicine, 491-494, 2015: [5] Ståhl F, Kolloch J, Almqvist H, Van Vlimmeren M, Soederman M, Falk Delgado A, Dual-layer detector cone-beam CT reduces artifacts and improves perception of intracranial structures, S2-SSPH02-4, RSNA 2022. [6] Wu P, Boone JM, Hernandez AM, Mahesh M, Siewerdsen JH. Theory, method, and test tools for determination of 3D MTF characteristics in cone-beam CT. Med Phys. 2021;48(6):2772-2789. [7] Siewerdsen JH, Cunningham IA, Jaffray DA. A framework for noisepower spectrum analysis of multidimensional images. Med Phys. 2002;29(11):2655-71. [8] Burgess AE, Wagner RF, Jennings RJ, Barlow HB. Efficiency of human visual signal discrimination. Science (1981);214(4516):93 94. [9] Burgess AE. Statistically defined backgrounds: performance of a modified nonprewhitening observer model. J Opt Soc Am A Opt Image Sci Vis. 1994;11(4):1237-42. [10] Eckstein M, Bartroff J, Abbey C, Whiting J, Bochud F. Automated computer evaluation and optimization of image compression of x-ray coronary angiograms for signal known exactly detection tasks. Opt Express. 2003;11(5):460-75. [10] ICRU (1995) Medical Imaging The Assessment of Image Quality Report 54. 7910 Woodmont Avenue, Bethesda. [11] ICRU (1995) Medical Imaging The Assessment of Image Quality Report 54. 7910 Woodmont Avenue, Bethesda [12] Burgess AE, Colborne B. Visual signal detection. IV. Observer inconsistency. J Opt Soc Am A. 1988;5(4):617-27 356
0
You can add this document to your study collection(s)
Sign in Available only to authorized usersYou can add this document to your saved list
Sign in Available only to authorized users(For complaints, use another form )