2.1

Data

The Quant CT-Training Phantom from Cascination^∗ was used in this work. It includes a liver with some lesions, a spine, the ribs, and a portal vein. The prior CT volume was acquired at 70 KVp with a Siemens SOMATOM X.cite scanner (Siemens Healthcare GmbH, Erlangen, Germany). The acquired image dimensions were 768 × 768×266 voxels with a voxel size of 0.479mm×0.479mm×1mm. The intraoperative CBCT volume was acquired using a Siemens ARTIS Icono C-arm scanner (Siemens Healthcare GmbH, Erlangen, Germany) at 87.5 KVp. The CBCT image dimensions were 512×512×368 voxels with a voxel size of 0.49mm×0.49mm×0.49mm. To mimic an interventional scenario, an ablation needle was inserted in the phantom before acquiring the intraoperative CBCT volume.

To simulate the intraoperative DTS projection images, the CBCT volume was forward-projected according to a newly proposed spherical ellipse scan geometry.¹ This protocol assumes that the source and the detector are mounted on a robotic C-arm and each moves along a spherical ellipse orbit. Fig. 1 illustrates the spherical ellipse acquisition setup. Mounted in opposite to the X-ray source, the detector performs an in-plane rotation in a way its rows are kept tangent to the scan trajectory. The detector is composed of 616 × 480 pixels with a 0.616 mm pixel pitch. The source to iso-center distance is set to 785 mm and the source to detector distance is set to 1200 mm. This trajectory is defined by two tomographic angles. In this work, we fixed the angles to ±15° and ±23° and simulated 72 projection views. The system acquisition has been implemented in the open-source Computed Tomography Library CTL^†.⁴

Figure 1.

Illustration of the spherical ellipse DTS scan trajectory.

2.2

Registration and Reconstruction

The proposed framework is illustrated in Fig. 2. To register the prior CT volume to the current intraoperative anatomy, an intermediate intraoperative DTS image is reconstructed from the intraoperative DTS projections. An iterative algebraic reconstruction technique (ART) with an ordered-subsets scheme⁵ is used. The relaxation parameter is estimated as in Ref. 6. At this stage the initial image voxels are set to zero. It was shown in Ref. 7 that the complexity of DTS resolution characteristics limits the accuracy of DTS-CT registration and a DTS-DTS registration is more precise. Therefore, from the prior CT volume we generated a set of reference digitally reconstructed radiographs (DRRs) using the same intraoperative DTS scan geometry described above and we reconstructed a reference DTS image. At this step the prior CT image was downsampled to 512 × 512 × 177 voxels with a voxel size of 0.7187mm × 0.7187mm × 1.5mm. Here it worth be noted that the reconstruction of the reference DTS image can be performed before the intervention as soon as the prior CT image is available. We performed the registration of the reference DTS image to the intraoperative DTS image using a rigid registration employing the Mattes mutual information as a similarity measure. We used the registration toolbox ANTsPy.⁸ The registration transform was then applied to the prior CT image. The intraoperative DTS projections are iteratively re-reconstructed using the co-registered prior CT as a first guess and including only a positivity constraint.

Figure 2.

Flowchart of the proposed reconstruction method.

2.3

Results

The registration results are shown in Fig. 3. One axial slice and one coronal slice where the needle tip and a liver lesion exist are shown in the upper row and the lower row respectively. Fig. 3(a) and 3(b) illustrate the prior CT volume and the intermediate intraoperative DTS volume (zero-initialized ART) respectively. The co-registered prior CT volume is shown in Fig. 3(c). For reference the same slice of the intraoperative CBCT volume is shown in Fig. 3(d). Despite that we cannot examine the difference image due to the difference in intensity values between the CT and the CBCT images, it is evident from the visual inspection that the prior CT volume is well registered to the intraoperative volume. The liver, the spine, and the lesions seem accurately aligned in (c) and (d).

Figure 3.

Registration results. (a) Prior CT, (b) intermediate intraoperative tomosynthesis, (c) coregistered prior CT, and (d) intraoperative CBCT. Upper row: one axial slice. Lower row: one coronal slice. Same slices are shown in (b), (c), and (d). The display window range is [-1000HU, 700HU].

Fig. 4 illustrates the reconstruction results. For comparison the same axial slice is shown in the upper row and the same coronal slice is shown in the lower row. Fig. 4(a) shows the DTS reconstruction with zero-initialized ART, without including the prior. Fig. 4(b) illustrates the reconstructed image with our proposed method using the co-registered prior CT as a starting image. For reference the intraoperative CBCT image is shown in Fig. 4(c). As expected, without including the prior, DTS exhibits poor resolution in the depth direction, the lesion is blurred and barely visible and the spine as well. A considerable improvement in image quality is observed with the prior-initialized reconstruction. It looks much similar to the intraoperative CBCT. These observations are further validated in the line profiles plotted in Fig. 5 corresponding to the vertical (graph on the left) and the horizontal (graph on the right) lines highlighted in yellow in Fig. 4(c).

Figure 4.

Reconstruction results. (a) zero-initialized ART, (b) proposed method (prior-initialized ART), and (c) intraoperative CBCT. Upper row: one axial slice. Lower row: one coronal slice. Same slices are shown in (a), (b), and (c). The display window range is [-1000HU, 700HU].

Figure 5.

Profile plots of the two lines highlighted in yellow in Fig. 4 (left: in the axial slice and right: in the coronal slice).

2.4

Discussion and Conclusion

In this work we proposed a prior-based iterative reconstruction framework to improve intraoperative digital tomosynthesis (DTS) reconstruction. We considered a high-quality prior CT volume is available prior to the intervention. We simulated intraoperative DTS projection images using a spherical ellipse DTS geometry.¹ The proposed framework is composed of two steps. First, the registration of the prior CT to the current anatomy, at this stage a reference prior DTS image simulated from the prior CT and an intermediate low-quality reconstructed intraoperative DTS image are co-registered, the registration transform is then applied to the prior CT. Second, an iterative ART reconstruction using the co-registered prior CT as a first guess is performed to further improve the intraoperative DTS image quality. The reference prior DTS can be reconstructed beforehand prior to the intervention as soon as the prior CT is available. While prior information have been incorporated in the iterative process itself in previous works, the initialization was mostly overlooked. In this work, we showed how important is the starting point of the iterative reconstruction in such a highly under-determined problem. A considerable improvement in the different structures visibility has been achieved. A still open question is how sensible is the reconstruction to insufficiently accurate registration. By the design of the method, information measured from the tomosynthesis trajectory are included in the prior-initialized ART and all missing information are inherited from the co-registered prior. In case of inaccuracies in the registration, this may lead to wrong conclusions while interpreting the images. Therefore, clinical experts will be included in the evaluation process in the future. In the current work, we only considered a rigid motion between the prior and intraoperative states, this is probably valid for head interventions, however in many other interventions, such as in the lungs, the abdomen, and the liver, theses organs are prone to elastic motions. Considering such complex motions is the subject of current works. Moreover, how low can we go with the angular range and the number of projections will be further investigated.