[DESCRIPTION]

5

FIELD OF THE INVENTION

The present invention relates generally to the correction of errors in the assumed acquisition geometry of a chest tomosynthesis X-ray

10 system, in order to reduce the occurrence of reconstruction image

artefacts during image reconstruction.

BACKGROUND OF THE INVENTION

15 In medical imaging, chest tomosynthesis (TS) is a form of limited

angle X-ray tomography with which 3D images of a patient can be computed. In chest tomosynthesis, projection images are acquired using a fixed flat panel detector, which is placed behind the

patient and a motorized X-ray tube, typically moving on a straight

20 line, to create projection images from a limited angle. An intrinsic characteristic of chest imaging and chest tomography is that the imaged body part -the chest- appears truncated' in the projection images. Nevertheless, other applications may be envisaged in

different clinical or non-clinical application domains, but showing

25 the same image acquisition constraints (namely truncation) .

Subsequently, a reconstruction algorithm is used to reconstruct section planes parallel to the detector. With slight modifications such as X-ray tube motorization and flat panel detector

30 synchronization, chest TS can be performed on standard X-ray

modalities. Chest TS offers a high in-plane resolution but suffers from a low depth resolution due to the limited acquisition angle of the projection images (as described in J. T. Dobbins, H. P. McAdams, D. J. Godfrey, and C. M. Li, "Digital tomosynthesis of the chest,"

35 Journal of Thoracic Imaging, vol. 23, no. 2, pp. 86-92, 2008.) . Nevertheless, chest TS is capable of separating overlapping anatomy into subsequent section planes, making the detection of certain pathologies easier compared to standard X-ray images (refer to N. Molk and E. Seeram, "Digital tomosynthesis of the chest: A

literature review," Radiography, vol. 21, no. 2, pp. 197-202,

2015. ) .

In order to reconstruct high quality TS images, accurate knowledge of the relative positions of the X-ray source and detector is crucial. Inaccuracies in the assumed acquisition geometry lead to reconstruction artefacts such as striping and blurring. For chest tomosynthesis, an important geometric parameter is the orientation of the detector relative to the linear motion path of the X-ray tube. Even inaccuracies smaller than 0,5° on the detector

orientation may lead to significant image reconstruction artefacts.

When acquisitions are performed with a rectangular detector in a table, these inaccuracies cannot be easily avoided as the mechanical detector assembly is allowed to be rotated 90° by the operator, resulting in inaccurate orientation due to mechanical ^" limitations of the locking mechanism of the assembly itself. Moreover in case of mobile tomosynthesis, the detector orientation is even more

difficult to control as there is no mechanical alignment or coupling between the X-ray source trajectory and the detector angle.

Therefore, methods are needed to accurately calibrate the

acquisition geometry. The acquisition geometry can be measured offline (meaning not during an acquisition) , using a calibration phantom (as for instance described by: H. Miao, X. Wu, H. Zhao, and H. Liu, "A phantom-based calibration method for digital x-ray tomosynthesis." J. Xray. Sci. Technol . , vol. 20, no. 1, pp. 17-29, jan 2012.), and online using the projection images.

For online calibration, radio opaque markers can be used to derive the acquisition geometry (see R. Han, L. Wang, Z. Liu, F. Sun, and F. Zhang, "A novel fully automatic scheme for fiducial marker-based alignment in electron tomography." J. Struct. Biol., vol. 192, no. 3, pp. 403-17, dec 2015., or see as well US patent US8000522), but this method is cumbersome since the markers need to be applied very accurately prior to each acquisition.

Another way is to exploit data consistency conditions, which describe redundancies between projection images. From these

conditions, a cost function can be formulated that, after

minimization, leads to the optimal geometric parameters such as detector orientation and position. Such a technique was developed for estimating and correcting the geometric parameters in a cone beam computed tomography setup, based on epipolar consistency conditions (ECC) (see C. Debbeler, N. Maass, M. Elter, F.

Dennerlein, and T. M. Buzug, "A New CT Rawdata Redundancy Measure applied to Automated Misalignment Correction," Proc. Fully Three- dimensional Image Reconstr. Radiol. Nucl . Med., pp. 264-267, 2013.). The same technique is disclosed in a patent application

WO2014/108237. However, improvements are necessary for tomosynthesis, as the ECC are not robust enough to deal with truncated projections. In recent scientific work, a heuristic weighting function was introduced to weigh the gray values in the projection images, depending on the fraction of the ray passing the part of the object that is visible on all projections and an estimation of the maximal object thickness (see A. Grulich, Tobias and Holub, Wolfgang and Hassler, Ulf and Aichert, Andre and Maier, "Geometric Adjustment of X-ray

Tomosynthesis," in Fully Three-Dimensional Image Reconstr. Radiol. Nucl. Med., Newport, 2015, pp. 468-471.). This described method thus describes a weighting method which relies on the estimated patient size (thickness) which is not convenient to provide good results.

SUMMARY OF THE INVENTION The present invention provides a method and system for chest tomosynthesis, comprising a method to improve the robustness of the calculation method used to correct for discrepancies between the assumed acquisition geometry and the actual acquisition geometry of a tomosynthesis image acquisition, as set out in independent claim 1 and 6. It should be noted that this invention is not limited to the application in chest tomosynthesis only, and that other applications may be envisaged in different clinical or non-clinical application domains, where the same image acquisition constraints are

encountered (namely truncation) .

The above-described aspects are solved by a system as set out in claim 1 to 11.

In the context of this invention, an acquisition geometry has to be understood as the geometric arrangement of the different components making up a digital imaging modality, being the geometric setup defined in terms of a set of geometric parameters of the arrangement of the components of a so-called image acquisition unit. In order to be functional, the acquisition geometry for a tomosynthesis imaging needs to be accurately known for the reconstruction algorithm to work properly. However, the real acquisition geometry may deviate slightly from the assumed acquisition geometry which is the

theoretical setup of the acquisition geometry. A data acquisition unit has to be understood as the configuration of components to perform an X-ray image acquisition, commonly named an X-ray

modality.

The components making up such a system are typically a patient positioning device, an X-ray source and a digital image detector. The patient positioning device is required to position the patient or object in a stable and accessible way (such as a table, or different positioning unit) for the X-ray source and digital image detector. The positioning device is however a recommended, but optional component for a chest tomosynthesis system; the patient may also be positioned while standing freely close to the digital image detector . The X-ray source typically is an X-ray tube mounted on a stand or mount, and is driven by an X-ray generator supplying the high voltage current to the tube filament at the moment of exposure. The digital image detector acquires the X-ray shadow of the imaged patient or object casted by the X-rays originating from the X-ray source. The projection image is acquired and converted into a digitally readable format. This digitally readable image is called in the context of this application a raw projection image as it comprises the raw or unprocessed image data from a single projection image acquisition.

The projection images of an object are acquired in different positions, meaning that in most cases the object is kept immobilized while either the X-ray source, or the digital image detector (or both) is moved to a different position before acquiring a subsequent projection image. In most cases, the X-ray detector and/or the image detector describe a geometrical trajectory such as a circular or linear path with respect to a fixed reference point in the

acquisition geometry or to each other. The acquisition geometry describes this relation between X-ray source, image detector and imaged object for subsequent acquisitions. In chest tomosynthesis acquisitions, mostly only the X-ray source is moved along a linear path on which at fixed intervals the movement of the X-ray source is interrupted and acquisitions are made of the imaged object. The total angle covered by the X-ray source between the first and last acquisition of as seen from the detector is between 5 and 15°.

A data processing unit subsequently performs various processing tasks on the projection images that have been transferred from the data acquisition unit. For example, the processing unit performs on the projection images desired operations including data sampling, data shifting, filtering, selection and reconstruction. The

processing unit calculates reconstructed images based on the projection images. As will be described later in greater detail, the processing unit performs certain filtering operations on the projection images before or after reconstruction. The functions of the data processing unit can be implemented on a computer system, and as such the present invention can be

implemented as a computer program product adapted to carry out the steps set out in the description. The computer executable program code adapted to carry out the steps set out in the description can be stored on a computer readable medium.

Specific examples and preferred embodiments are set out in the dependent claims.

The present invention is beneficial in that the correction method is capable to correct for discrepancies between the assumed and actual acquisition geometry during reconstruction of a truncated

tomosynthesis image data set, and this without the need for upfront calibration of the tomosynthesis X-ray system. The correction method improves the robustness of the reconstruction method, and prevents the appearance of reconstruction artefacts in the resulting

reconstructed image.

Further advantages and embodiments of the present invention will become apparent from the following description and drawings . BRIEF DESCRIPTION OF THE DRAWINGS

Fig. 1 illustrates the geometric setup of a chest tomosynthesis acquisition in one embodiment of the invention. N X-ray images are acquired on a stationary flat panel Θ with an X-ray tube that moves on a straight line S. The angle μ represents the orientation of the digital image detector relative to the X-ray source path. Most chest tomosynthesis systems are designed so that μ = 0.

Fig. 2 shows the shapes of different proposed weighting functions which may be applied to the projection image prior to or after calculating the redundant cost planes function C _RP . Fig. 3 shows the shape of cost function for estimated digital image detector angle μ without truncation filtering applied first. The cost function is calculated for a simulated chest tomosynthesis exam using the XCAT phantom.

Fig. 4 shows the shape of different cost functions calculated for the same simulated chest tomosynthesis exam. Vertical Gaussian truncation filters where used with varying values of w = Li - L ₂ . Fig. 5 shows the shape of different cost functions calculated for the same simulated chest tomosynthesis exam. Horizontal Gaussian truncation filters where used with varying values of w = Li = L ₂ .

Fig. 6 shows the shape of different cost functions calculated for the same simulated chest tomosynthesis exam. Gaussian truncation filters where used in both horizontal and vertical direction. The calculated results for the disclosed experimental setup show that the cost function reaches a minimum around an angle correction of about 13° in case that filters with w<0.01 are chosen. The correct result (around 10°) prominently appears as the minima of the different cost functions calculated with values w>0.05.

DETAILED DESCRIPTION OF THE INVENTION In the following detailed description, reference is made in sufficient detail to the above referenced drawings, allowing those skilled in the art to practice the embodiments explained below.

Embodiments of the present invention provide a system and a method for acquiring and reconstructing a chest tomosynthesis projection image data set. The method disclosed by this invention is based on the original method of Debbeler (already referred to above) , and is a modification of this method to improve the robustness against truncation. A schematic representation of chest tomosynthesis is illustrated in Fig. 1. Projections are acquired of a patient f with a stationary detector Θ. The N subsequent positions of the X-ray source are indicated as s {X±) . In one embodiment of the invention, the X-ray source moves on a linear path S. In other embodiments of this invention, the source might move on a circular path over a limited angle .

The detailed description given here will further focus on the first embodiment in which the source moves on a linear path. The

coordinate frame attached to the detector has axes (U, V, W) .

The epipolar redundancy criterion describes the following

relationship between two projection images. For two source points s( An) and s{X _n -), multiple planes Ω can be drawn that intersect both points and the detector along an intersection which can be

parameterized by an angle μ and a distance 2 from the detector center (u ₀ ,v ₀ ) . For noiseless acquisitions and without truncation, it can be proven that

8 (λ _η ,μ,1) = _ξ ^(λ _η .,μ,1) (i)

with g ₃ defined as: g ₃ ( „, ,l) =—g _{2 B} , ,l) (2)

01

where g ₂ {λ _η , , /) = J g _[ (λ _η ,I cosμ - 1 sin μ, I sin μ + 1 cosμ)ώ ( 3 ) and g _l ( _n ,u,v) = (4)

with p the projection data, w the normal of the detector and t(n,u,v) the direction of the ray arriving in detector pixel (u,v) of the n ^th projection. Based on this, a redundant cost planes function C _RP can be derived which reaches a minimum if the geometric parameters (and hence A and 2) are correctly estimated:

„=0 / 2/=-I _max

Although that other geometric parameters may be optimized to achieve a minimum in the cost planes function , a first embodiment of this invention optimizes the detector orientation μ, relative to the linear motion path of the X-ray tube.

More specifically in chest tomosynthesis, substantial truncation of the object is present in the projections both in the horizontal and vertical direction. In the case of even very small angle deviations μ it can be expected that the largest inconsistencies between the projections are found at the top and bottom regions of the

projections, as certain parts of the patient will not be imaged in this case, depending on the acquisition angle of the tube.

Intuitively, the horizontal truncation would cause fewer

inconsistencies. While in the prior art it has been attempted to optimize the efficiency of the cost function calculation by applying a weighting function to reduce the weight in the cost function of pixels that were suspected to contain information that was not present in all projections (see A. Grulich, Tobias and Holub,

Wolfgang and Hassler, Ulf and Aichert, Andre and Maier, "Geometric Adjustment of X-ray Tomosynthesis," in Fully Three-Dimensional Image Reconstr. Radiol. Nucl . Med., Newport, 2015, pp. 468-471.), the invention described in this application proposes a different approach.

The equation above (3) is very sensitive for pixels near the upright (horizontal) image edges, even if an object would have been imaged that fitted perfectly on the detector without horizontal truncation. A small deviation from 0 in μ would cause a large part of the image pixels in the image border to fall off the intersection with the plane Ω, causing a large discontinuity in g ₃ and thus making the entire cost function C _BP unstable. Therefore, in order to reduce the impact of pixels near the edge of the projections on the cost function and in a specific embodiment of this invention, a specific weighting function is applied. In the regions of relative width Li and Ri, weights are increased from 0 to 1 according to a Gaussian distribution with Li = 3σ and respectively i?i = 3σ. See Fig 2.

In another embodiment, the same filter is also disclosed to

compensate for vertical truncation.

In yet other embodiments, different shapes of weighting curves are proposed, such as linear, rectangular or sigmoid shaped sighting curves (such as depicted in Fig 2) . The above has been confirmed by a simulated chest tomosynthesis exam experiment using the XCAT phantom as the imaged object. A set of 11 simulated projections was generated using a mathematical (computer) toolkit with a source-image distance of 120cm and linear tube motion path of 20cm. Detector size was set to 360 x 420 pixels of 1 mm size. The detector was placed at a relative rotation of 10° with the motion path of the X-ray tube. Experiments were performed to estimate this simulated detector rotation, through the calculation of the said redundant cost planes functions. The maximum achievable accuracy of the estimation of the detector rotation is related to the detector size; the maximum accuracy Δ is defined as the angle increment for which a rays passes through a neighbouring pixel at the edge of the image: Δ = tan ^-1 (1/210) = 0,27° .

Fig 3 clearly shows that the estimation of the orientation angle, based on the uncorrected raw projection images results in the estimation of μ = 12,1° as the correct angle. As can be seen in Fig 3, the cost function C _RP achieved a minimum value at 12,7° which is however well above the ⁽ true' angle of 10°. Fig 4 and 5 show similar calculations of the detector rotation, but here the different cost function curves are calculated based on the same set of projection images, but where respectively vertical and horizontal Gaussian truncation filters with varying values of w = L ₂ = L ₂ where applied before the reconstruction. Surprisingly, when the applied filter widths are w > 0.10, the minimum of C _RP moves towards 10°. When w is further increased beyond 0.40, the calculations show that the value of μ no longer is estimated correctly, which is not a surprise since the amount of useable image data is dramatically reduced and potentially no longer representative. Combining both horizontal and vertical filters (Fig 6) more or less produces the same results, confirming the need for horizontal filtering despite the mainly vertical truncation inconsistencies.

In another embodiment, the same weighting functions are applied, but as opposed to the previous embodiment, to the calculated cost functions themselves (without applying the weighting functions to the projection images) . This method is supported by the fact that the calculation of the weighting functions out of the different projection images is a mathematically linear operation.