Next generation tomosynthesis image acquisition optimization for dedicated PET-DBT attenuation corrections

Abstract

A next generation tomosynthesis (NGT) prototype is under development to investigate alternative acquisition geometries for digital breast tomosynthesis (DBT). A positron emission tomography (PET) device will be integrated into the NGT prototype to facilitate DBT acquisition followed immediately by PET acquisition (PET-DBT). The aim of this study was to identify custom acquisition geometries that (1) improve dense/adipose tissue classification and (2) improve breast outline segmentation. Our lab’s virtual clinical trial framework (OpenVCT) was used to simulate various NGT acquisitions of anthropomorphic breast phantoms. Five custom acquisition geometries of the NGT prototype, with posteroanterior (PA) x-ray source motion ranging from 40–200 mm in 40 mm steps, were simulated for five phantoms. These acquisition geometries were compared against the simulation of a conventional DBT acquisition geometry. Signal in the reconstruction was compared against the ground truth on a voxel-by-voxel basis. The segmentation of breast from air is performed during reconstruction. Within the breast, we use a threshold-based classification of glandular tissue. The threshold was varied to produce a receiver operating characteristic (ROC) curve, representing the proportion of true fibroglandular classification as a function of the proportion of false fibroglandular classification at each threshold. The area under the ROC curve (AUC) was the figure-of-merit used to quantify adipose-glandular classification performance. Reconstructed breast volume estimation and sensitivity index (d′) were calculated for all image reconstructions. Volume overestimation is highest for conventional DBT and decreases with increasing PA source motion. AUC and d′ increase with increasing PA source motion. These results suggest that NGT can improve PET-DBT attenuation corrections over conventional DBT.

1. INTRODUCTION

Digital breast tomosynthesis (DBT) has been shown to improve breast cancer detection and reduce patient callbacks,¹ indicating an increase in sensitivity and specificity.² Despite these advantages over full-field digital mammography, conventional DBT does not favor the detection and characterization of calcifications; and lacks prognostic capability.² Next generation tomosynthesis (NGT) is under development to investigate potential advances in breast cancer diagnostics and prognostics.

The first stage of development for the NGT prototype was the incorporation of novel x-ray acquisition geometries, including two-dimensional (2D) x-ray source motion in the mediolateral (ML) and posteroanterior (PA) directions and craniocaudal (CC) detector motion. We have evaluated the imaging capabilities of the preliminary NGT prototype using tests of physics and image quality.³ NGT has been shown to support isotropic and high-quality super resolution,^4,5 improved breast volume estimation,^6,7 and reduced image reconstruction artifacts.^6,8

Two planar positron emission tomography (PET) detectors will be integrated with the NGT prototype to provide functional imaging as a dedicated PET-DBT device (Figure 1). As an emission imaging modality, the 511 keV annihilation photons that form a PET signal are susceptible to Compton scatter and photoelectric absorption within the breast prior to detection leading to image nonuniformities. In order to achieve quantitative PET images, accurate attenuation correction of the PET data is necessary. In commercial whole-body PET/CT, the CT image is routinely used for PET attenuation correction. In a similar fashion, we anticipate using the DBT images for accurate attenuation correction of the PET data.

Figure 1:

Design and photograph of the PET-DBT prototype. During DBT Mode, x-ray images are acquired while PET detectors are retracted into a homing position. Immediately following a tomosynthesis acquisition, PET detectors are positioned above and under the compressed breast and a PET scan is acquired (PET mode).

Adipose and fibroglandular tissue are the primary breast tissues in the female breast.⁹ The task of determining the 3D volumetric outline (breast segmentation and breast volume estimation) and the task of classifying the adipose and fibroglandular tissue using DBT image reconstructions are two of the primary concerns for attenuation correction of the PET-DBT device.¹⁰ DBT is an under-sampled tomographic technique and is thus prone to cone-beam artifacts and out-of-plane reconstruction artifacts in the depth dimension due to overlapping structures. Cone-beam artifacts overestimate the breast volume and out-of-plane artifacts introduce difficulties for distinguishing fibroglandular tissue from adipose tissue. For these reasons, it is especially difficult to determine the true percent density of patient breasts without known ground truth.¹¹ The 2D x-ray source motion of the NGT prototype can be optimized using a virtual clinical trial (VCT) framework for these tasks with known ground truth. We hypothesize introducing PA source motion will improve breast volume segmentation and adipose-fibroglandular tissue classification.

2. MATERIALS & METHOD

Image Acquisition

We used our open-source virtual clinical trial framework (OpenVCT)¹² to evaluate breast segmentation (breast volume estimation), and adipose-glandular classification for custom NGT acquisition geometries compared with conventional DBT. Five phantoms were generated and simulated using mediolateral (ML) compression (Figure 3). Each phantom consisted of a random distribution of adipose and glandular tissue compartments with skin surrounding these tissues. Phantoms were created using an isotropic voxel resolution of 0.1 mm. Each phantom has a volume of 700 mL and measures 7.8 cm × 6.3 cm × 20.5 cm after compression.

Figure 3:

The central slice and maximum intensity projection render of each of the five phantoms are shown. Adipose, skin, and glandular breast tissue are the only materials that were simulated for the phantoms used in this study.

Five bowtie acquisition geometries that incorporate increasing distances of x-ray source motion in the PA direction, forming a V (Figure 2), were simulated for each phantom. In addition, a conventional acquisition geometry with the same number of projections and no PA source motion, was used to acquire projection images for all 5 phantoms. Each acquisition geometry consisted of 15 x-ray projections. The source to detector distance was constant at 652 mm. Images were acquired using 35 kVp, 70 mAs (4.67 mAs/projection), and aluminum filtration per projection.

Figure 2:

The bowtie acquisition geometries that were used for this study. The black dots represent the common projection locations amongst the 5 geometries. The maximum extent of PA-source motion (y) is indicated in the legend for each geometry. The coordinate (0,0) represents the origin of the NGT prototype at the center of the chest wall.

Simulated projection images were used to create image reconstructions of each acquisition using commercial reconstruction software. The outline of the breast was segmented from the background for each image reconstruction. Reconstructions were produced with a voxel resolution of 0.1 × 0.1 × 0.5 mm³ and processed with contrast enhancement filtering isotropically and bilaterally on each projection image (Piccolo version 4.0.5, Real Time Tomography, Villanova, PA).

Quantitative assessment

The reconstructed breast volume was estimated using the full-resolution image reconstructions. The reconstruction mask of each simulated acquisition was determined by the total volume of voxels with a signal greater than zero. Then, the reconstruction mask was normalized to the phantom mask of non-air voxels for each respective phantom (ground truth).

The reconstruction of the adipose-fibroglandular structures was assessed using a binary classification metric akin to ROC methods previously developed in our lab.⁶ This analysis is computationally burdensome for full-resolution phantoms and reconstructions, so both the phantom and image reconstructions were down sampled to an isotropic resolution of 0.5 mm. The phantoms were down-sampled using nearest-neighbor interpolation (mode), and image reconstructions were down sampled using bilinear interpolation. The down-sampled image reconstructions were compared against the phantom ground truth.

ROC curves were produced by using the signal intensity of the image reconstruction as a threshold for a binary classification task. Signals above the threshold were classified as glandular tissue (including voxels containing skin) and signals below were classified as adipose. The true positive rate (TPR) was graphed against the false positive rate (FPR) for each image reconstruction as a function of the threshold value. Area under the curve (AUC) was calculated for each ROC. Colormap images, showing voxel-by-voxel representations of the binary classifications, were created to visualize the differences in image reconstructions of the various acquisition geometries.

In binary classification tasks, the sensitivity index (d′) quantifies the effective signal-to-noise ratio of a two-alternative forced choice classification.¹³ We use d′ to determine differences between different geometries. d′ is obtained using an inverse error function and AUC:

$$d^{\prime }=2\cdot {\text{erf}}^{-1}(2\cdot \text{AUC}-1)$$ (1)

We then calculated the difference in d′ (Δ_d′) between bowtie acquisition geometries (b) and the conventional acquisition geometry (c) to quantify performance:

$${\Delta }_{d^{\prime }}=\frac{{d^{\prime }}_b-{d^{\prime }}_c}{{d^{\prime }}_c}$$ (2)

3. RESULTS AND DISCUSSION

The central slice of the bowtie acquisition geometry with 80 mm of PA source motion is shown as an example for all five phantoms in Figure 4. Image reconstructions for the conventional and remaining bowtie acquisition geometries have a similar overall appearance. It is difficult to discern any differences qualitatively between the reconstructions of the various acquisition geometries in the conventionally reconstructed sagittal slices. Therefore, coronal and transverse reconstruction slices near respective midplanes of Phantom 2 are shown for the six geometries alongside the ground truth in Figure 5.

Figure 4:

Examples of the central slice of image reconstructions for each of the five phantoms in the conventionally reconstructed sagittal slices (bowtie acquisition with 80 mm PA source motion).

Figure 5:

Coronal and transverse reconstruction slices of the conventional acquisition geometry (0 mm) and the five NGT acquisition geometries with increasing PA source motion (40 – 200 mm).

Coronal and sagittal reconstruction slices show trends of improved breast outline delineation and signal contrast of overlapping tissue that scale with PA source motion. Coronal slices Additionally, the effect of cone-beam artifacts is observed in the transverse slices. The artifacts are most prominent in the conventional acquisition geometry (0 mm) and are reduced for acquisition geometries with PA source motion greater than 80 mm.

An example of the binary classification is shown for the central slice of a conventional acquisition geometry and an NGT acquisition geometry with 200 mm of PA source motion as a colormap in Figure 6 (sagittal plane) and Figure 7 (transverse plane). This example is shown with a signal threshold at the optimal operating cut-point and was determined by the minimum Euclidean distance from the top left corner (0,1) of the ROC curve. The colormap images of the NGT geometry show fewer out-of-plane reconstruction artifacts compared with the conventional geometry.

Figure 6:

Binary classification colormap of a conventional reconstruction and NGT acquisition with 200 mm of PA source motion. The classification image shows true positives (TP), true negatives (TN), false positives (FP), and false negatives (FN). For the Ground Truth and Reconstruction images, white represents fibroglandular tissue.

Figure 7:

Coronal slice examples of the colormap images. The NGT acquisition geometry shows fewer out-of-plane artifacts (false positives) compared with the conventional geometry.

The mean volume overestimation of the five phantoms is graphed against PA source motion in Figure 8 (left). The conventional DBT acquisition geometry shows the highest overall volume estimation at 112.4% of the ground truth. Volume estimation decreases with increasing PA source motion to a low of 108.0% with 200 mm of PA source motion. This result is consistent with previous results^6,7 and indicates that PA source motion in DBT decreases artifacts in image reconstructions.

Figure 8:

Mean volume overestimation as a function of PA source motion (left). Improvement in d′ for bowtie acquisition geometries with PA-source motion ranging from 40 mm to 200 mm (right).

Average improvement in d′ is graphed against PA source motion in Figure 8 (right). Positive values indicate that all bowtie acquisition geometries show improvement in d′ over conventional DBT. The greatest overall improvement is 8% for two of the five phantoms. These results are preliminary and need to be evaluated further. Phantoms of different volumes and various percent density will be analyzed to evaluate additional factors that can affect volume overestimation and the sensitivity index. The impact of image processing has not been evaluated. For the sake of this study, image processing was the same across all acquisition geometries. Acquisition-geometry specific image processing could provide improved filtering of out-of-plane artifacts for novel geometries. We chose the simplest method of segmentation – threshold segmentation – to test the physics of the NGT system and optimize image acquisition geometries. This approach will improve input data for the ultimate segmentation task.

4. CONCLUSION

The PA source motion of the custom NGT acquisition geometries has been shown to improve volume estimation and adipose-glandular classification for DBT using a simple threshold-based method. Results suggest that the NGT acquisition geometries can improve the accuracy of PET-DBT attenuation corrections. This method will be used to evaluate additional phantom parameters and acquisition geometries. These results can also be used to help determine a superior classification or tissue segmentation method for DBT.