Breast cancer is the leading cause of cancer-related death among women.
It is also difficult to diagnose.
Nearly one in 10 cancers is misdiagnosed as not cancerous, meaning that a patient can lose critical treatment time.
On the other hand, the more mammograms a woman has, the more likely it is she will see a false positive result.
After 10 years of annual mammograms, roughly two out of three patients who do not have cancer will be told that they do and be subjected to an invasive intervention, most likely a biopsy.
Breast ultrasound elastography is an emerging imaging technique that provides information about a potential breast lesion by evaluating its stiffness in a non-invasive way.
Using more precise information about the characteristics of a cancerous versus non-cancerous breast lesion, this methodology has demonstrated more accuracy compared to traditional modes of imaging.
At the crux of this procedure, however, is a complex computational problem that can be time-consuming and cumbersome to solve.
But what if instead we relied on the guidance of an algorithm?
Assad Oberai, USC Viterbi School of Engineering Hughes Professor in the Department of Aerospace and Mechanical Engineering, asked this exact question in the research paper, “Circumventing the solution of inverse problems in mechanics through deep learning: application to elasticity imaging,” published in Computer Methods in Applied Mechanics and Engineering.
Along with a team of researchers, including USC Viterbi Ph.D student Dhruv Patel, Oberai specifically considered the following: Can you train a machine to interpret real-world images using synthetic data and streamline the steps to diagnosis?
The answer, Oberai says, is most likely yes.
In the case of breast ultrasound elastography, once an image of the affected area is taken, the image is analyzed to determine displacements inside the tissue.
Using this data and the physical laws of mechanics, the spatial distribution of mechanical properties – like its stiffness – is determined.
After this, one has to identify and quantify the appropriate features from the distribution, ultimately leading to a classification of the tumor as malignant or benign.
The problem is the final two steps are computationally complex and inherently challenging.
In the research, Oberai sought to determine if they could skip the most complicated steps of this workflow entirely.
Cancerous breast tissue has two key properties: heterogeneity, which means some areas are soft and some are firm, and non-linear elasticity, which means the fibers offer a lot of resistance when pulled instead of the initial give associated with benign tumors.
Knowing this, Oberai created physics-based models that showed varying levels of these key properties.
He then used thousands of data inputs derived from these models in order to train the machine learning algorithm.
Synthetic Versus Real-World Data
But why would you use synthetically-derived data to train the algorithm?
Wouldn’t real data be better?
“If you had enough data available, you wouldn’t,” said Oberai. “But in the case of medical imaging, you’re lucky if you have 1,000 images.
In situations like this where data is scarce, these kinds of techniques become important.”
Oberai and his team used about 12,000 synthetic images to train their machine learning algorithm.
This process is similar in many ways to how photo identification software works, learning through repeated inputs how to recognize a particular person in an image, or how our brain learns to classify a cat versus a dog.
Through enough examples, the algorithm is able to glean different features inherent to a benign tumor versus a malignant tumor and make the correct determination.
Oberai and his team achieved nearly 100 percent classification accuracy on other synthetic images.
Once the algorithm was trained, they tested it on real-world images to determine how accurate it could be in providing a diagnosis, measuring these results against biopsy-confirmed diagnoses associated with these images.
“We had about an 80 percent accuracy rate. Next, we continue to refine the algorithm by using more real-world images as inputs,” Oberai said.
Changing How Diagnoses are Made
There are two prevailing points that make machine learning an important tool in advancing the landscape for cancer detection and diagnosis.
First, machine learning algorithms can detect patterns that might be opaque to humans. Through manipulation of many such patterns, the algorithm can produce an accurate diagnosis.
Secondly, machine learning offers a chance to reduce operator-to-operator error.
So then, would this replace a radiologist’s role in determining diagnosis?
Definitely not. Oberai does not foresee an algorithm that serves as a sole arbiter of cancer diagnosis, but instead, a tool that helps guide radiologists to more accurate conclusions.
“The general consensus is these types of algorithms have a significant role to play, including from imaging professionals whom it will impact the most.
However, these algorithms will be most useful when they do not serve as black boxes,” said Oberai.
“What did it see that led it to the final conclusion?
The algorithm must be explainable for it to work as intended.”
However, these algorithms will be most useful when they do not serve as black boxes,” said Oberai.
“What did it see that led it to the final conclusion?
The algorithm must be explainable for it to work as intended.”
Adapting the Algorithm for Other Cancers
Because cancer causes different types of changes in the tissue it impacts, the presence of cancer in a tissue can ultimately lead to a change in its physical properties, for example a change in density or porosity.
These changes are can be discerned as a signal in medical images.
The role of the machine learning algorithm is to pick out this signal and use it to determine whether a given tissue that is being imaged is cancerous.
Using these ideas, Oberai and his team are working with Vinay Duddalwar, professor of clinical radiology at the Keck School of Medicine of USC, to better diagnose renal cancer through contrast enhanced CT images.
Using the principles identified in training the machine learning algorithm for breast cancer diagnosis, they are looking to train the algorithm on other features that might be prominently displayed in renal cancer cases, such as changes in tissue that reflect cancer-specific changes in a patient’s microvasculature, the network of microvessels that help distribute blood within tissues.
The use of palpation to determine the stiffness of a lesion has been used since the time of the ancient Greeks and Egyptians [1]. Stiff, non-mobile lesions of the breast have a high probability of being malignant. In vitro experiments it has been shown that the difference in stiffness of malignant breast lesions and benign breast lesions is substantial with little overlap [2]. These characteristics of malignant breast masses suggest that elastography should have a high accuracy for characterization of breast lesions.
Both strain elastography (SE) and shear wave elastography (SWE) [3] have been used to evaluate breast pathology with high sensitivity and specificity. For reasons not fully understood malignant breast lesions appear larger on elastography compared to B-mode while benign lesions appear smaller. This has allowed for a semi-quantitative method in SE to characterize breast lesions. SWE can determine the lesion stiffness and determine with high probability of a lesion is benign or malignant. On SWE, some cancers have properties that do not allow for good shear wave propagation and may be interpreted as false negative lesions. This has been overcome by the addition of a quality measure (QM) which can detect the quality of the shear waves and alert the interpreter of an erroneous measurement.
Breast elastography has now been clinically available for over 10 years and has continued to improve. We now understand its advantages and limitations. The techniques have not been widely accepted due to a substantial learning curve for performing and interpreting examinations. Improvements in the technology are being developed to overcome these limitations and may make elastography a primary method of characterization of breast lesions.Go to:
Types of Elastography
There are two types of ultrasound elastography strain (SE) and SWE [3].
SE is a qualitative technique. This technique evaluates the changes in tissues when an external force is applied. Softer tissues deform more than stiffer tissues.
There are two methods of applying the stress, manual compression and release or using an acoustic radiation force impulse (ARFI) push pulse.
Most vendors offer manual compression strain on their system.
The systems vary significantly in the degree of stress that is needed for optimal images. Some systems require minimal compression and release while others require a moderate amount of compression and release.
There is a learning curve for each system.
There is a sweet spot where the algorithm is most accurate.
Too little or too much compression will lead to suboptimal elastograms (Fig. 1).
Most systems have a visual display of the amount of compression release applied so the operator can find the optimal technique.
When using the ARFI technique for SE, which is presently only offered by one vendor, the transducer is held still with minimal pressure on the breast and the ARFI pulse initiated.
The learning curve for the ARFI technique is less than that of the compression-release techniques.
The systems that require minimal compression and release are easier to learn as most require just the patient breathing or heart beat to generate the compression and release.
The operator just needs to hold the transducer with minimal compression on the breast.
Those that require more compression and release have a more significant learning curve and have more artifacts particularly in the near field.
SWE techniques use the ARFI technology to generate shear waves that propagate perpendicular to the ARFI pulse.
Therefore, ARFI technology can be used to generate both SE, measuring the axial displacement, and SWE measuring the speed of the perpendicular shear waves generated.
It is important to remember that the shear waves are tracked using B-mode ultrasound so optimal B-mode images are required to obtain accurate stiffness values.
The SWE technique is quantitative providing a stiffness value either by the shear wave speed in meters per second (m/sec) or converting to the Young’s modulus in kiloPascals [4].
The learning curve for SWE is less than that of SE although one important factor that affects both SE and SWE is pre-compression which is discussed in detail below [5,6].
Technique
One of the most critical factors in performing breast elastography is pre-compression [7]. When a material is compressed its stiffness increases (Fig. 2).
To perform accurate breast elastography, both strain and shear wave, minimal compression of the breast must be applied.
Usually some pre-compression is used in B-mode imaging as this straightens Cooper’s ligaments decreasing artifacts. In SE when significant pre-compression is applied the result is noise.
While with mild to moderate pre-compression there may be alternating good images and noise.
The good images are obtained in the upstroke of the compression cycle.
If serial images are not similar when performing SE, the reason is either too much pre-compression or movement of the lesion in the scanning plane. A method for applying minimal pre-compression with consistency has been proposed [7].
Another important factor to obtain adequate SE images is the same imaging plane of the lesion must remain in the field of view (FOV) throughout the data acquisition.
The patient should be positioned so that the transducer is perpendicular to the floor and the patient rotated so that the patient’s breathing moves the lesion within the image plane [8].
In SE, only the stiffnesses of tissue relative to each other are obtained.
Therefore, it is important to have several tissues present in the FOV. For breast SE, fat (softest tissue), fibro-glandular tissue, the pectoralis muscle and the lesion should be included.
In general, if the lesion is benign it will be of similar stiffness to the fibro-glandular tissues. Malignant lesions will be stiffer than all other tissues. Several color maps can be used.
It is the authors opinion that the grey scale map is the easiest to detect subtle changes between tissues and identify noise.
It is important recognize the color map used as some display red as stiff and other blue as stiff.
SWE provides a quantitative estimate of the tissue stiffness. For breast SWE the transducer is placed on the breast with minimal pre-compression and held still over the area of concern to obtain the measurement.
Both point shear wave elastography and two-dimensional (2D)-SWE have been used to evaluate breast lesions [9].
Breast masses, especially malignancies, tend to be very heterogeneous in stiffness. Therefore, the 2D-SWE technique is preferred as the larger FOV can depict the differences in stiffness and the area of highest stiffness can be identified. Table 1 lists the important technical factors when obtaining SE and SWE breast images.
Table 1.
Technical factors for performing breast elastography
Strain elastography | Shear wave elastography |
---|---|
Avoid pre-compression | Avoid pre-compression |
Use optimal compression/release for system | Hold transducer still |
No out of plane motion | No out of plane motion |
Position patient so respiratory motion is in plane with imaging | Position patient so lesion is at least 5 mm deep and less than 4 cm deep |
Have a large FOV containing fat, glandular tissue, pectoralis muscle and the lesion | Monitor the stiffness of fat to confirm no pre-compression |
FOV, field of view.
Interpretation of SE
For reasons not fully understood, breast cancers appear larger on the elastogram than on the B-mode image while benign lesions appear smaller on the elastogram than the B-mode image (Fig. 3).
This appears to be unique to breast imaging. Since SE is displaying relative stiffness differences, evaluating only the “color” on the elastogram does not provide adequate information for high sensitivity and specificity of characterizing breast masses. However, the size changes that occur are extremely sensitive and specific for characterizing breast masses.
There are three methods that have been proposed to interpret SE images of the breast: (1) elastogram to B-mode length ratio (E/B ratio); (2) 5-point color scale; and (3) comparing the stiffness of the lesion to the stiffness of fat (strain ratio).
Future Developments
Breast elastography has very high sensitivities and specificities for characterization of breast lesions. However, the learning curve of SE has limited its clinical acceptance. With SWE the learning curve is less steep but the problem of poor or no shear wave propagation in breast cancers is of concern due to possible false negative results. The development of a real-time stress map may solve some of these problems. This real-time technique displays a map which depicts the degree of applied stress. The technique will depict when uneven stress is applied (Fig. 13). It can guide where the appropriate ROIs should be placed for strain ratio.
Three-dimensional (3D) SWE has been available for some time and 3D strain may be clinically available in the near future [32,33].
It is unclear if 3D will provide improved results over 2D SWE or 2D SE. 3D does allow for visualization of the lesion in the coronal plane, ability to view the stiffness of the entire lesion, and may allow for whole breast screening.
Other new techniques are becoming clinically available including vibratory SWE where a mechanical device is used to create standing shear waves which are detected with B-mode ultrasound [34].
This technique also allows for varying the frequency of the mechanical stress possible allowing for additional diagnostic information.
Opto-acoustic imaging (OA) is another technique being used to characterize breast lesions. In a large study of 1,690 patients with 1,757 masses of which 1,079 (61.4%) were benign and 678 (38.6%) were malignant, OA was able to downgrade 40.8% of benign masses [35].
The LRN was 0.094 indicates a negative examination can reduce a pre-test probability of 17.8% (low BI-RADS 4B) to a post-test probability of 2% (BI-RADS 3).
Provided by University of Southern California