Visualization of Lesions in Breast MRI

From MeliWiki
Jump to: navigation, search

Motivation and Problem Description

Breast cancer is the most common cancer among women, but has an encouraging cure rate if diagnosed in an early stage. Magnetic resonance (MR) imaging is an emerging and promising new modality for detection and further evaluation of clinically, mammographically and sonographically occult cancers.

Acquisition of temporal sequences of between three to seven MR images depicting the kinetics of contrast agent molecules in the breast tissue allow for detecting and assessing suspicious tissue disorders with high sensitivity, even in the mammographically dense breasts of young women. Yet, the multitemporal nature of the three-dimensional image data poses new challenges to radiologists as the key-information is only perceivable if all images of the temporal sequence are considered simultaneously. Additionally, radiologists need techniques that allow them not to only analyze a single tumor at a time, but enable them to explore many different tumor lesions at the same time. Concurrent tumor inspection is not feasible since it is inter- and intra-observer dependent and the visualization of the entire signal space is limited by the high-dimensionality of the signal space. In MRI, multiple 3D T1-weighted MR images of both breasts are acquired over a period of 5-9 min while a contrast agent (CA) passes through the tissue. A typical sequence of images consists of one precontrast image acquired before injection of a CA bolus and a series of postcontrast images recorded afterwards over.1 Thus, a time-series signal, i.e., a vector reflecting the local signal intensities at the time points of image acquisition, is associated with each voxel. Due to characteristic changes in the structure of benign and malignant tissue influencing the flux of CA molecules between the blood pool and tissue, characteristic timeseries signals can be observed for different tissue types. Interpretation of these time-series signals allows for detecting cancer with high sensitivity, even in the radioopaque breast of young women, as well as for assessing the type of disorders in a non-invasive fashion [1].


Spatio-Temporal Classification

However, while the presence of a suspicious tissue disorder can already be identified by means of a strong signal enhancement in an early postcontrast image, the course of the entire timeseries signal has to be considered for differentiating benign and malignant tissue. The computer assisted interpretation of time-series signals as measured during a DCE-MRI examination for each image voxel represents one of the major steps in designing CAD systems for breast MRI. Kuhl et al. have shown that the shape of the time-series signals represents an important criterion in differentiating benign and malignant masses [2]. The results indicate that the enhancement kinetics, as represented by the time-series signals visualized in Fig. 2, differ significantly for benign and malignant enhancing lesions and thus represent a basis for differential diagnosis: plateau or washout-time courses (type II or III) prevail in cancerous tissue. Steadily progressive signal intensity time courses (type I) are exhibited by benign enhancing lesions, albeit these enhancement kinetics are shared not only by benign tumors but also by fibrocystic changes.

Figure 1: Schematic drawing of the three types of time-series signals according to [2]. Type I corresponds to a straight (Ia) or curved (Ib) line; enhancement continues over the entire dynamic study. Type II is a plateau curve with a sharp bend after the initial upstroke. Type III is a washout-time course. For visual inspection, the time-series signals are typically displayed as relative-enhancement curves depicting the signal enhancement with respect to the signal intensity in the precontrast image.

Even though the time-series signals enable radiologists to infer information about the tissue state, assessing the signal characteristics is a time-consuming task which needs experience and expertise. It becomes further complicated due to the heterogeneity of lesion tissue causing the signal characteristics to vary spatially. Also these spatial variations of signal characteristics reflect specific tissue properties and should be taken into account for assessing the state of lesions. Different computerized approaches have been proposed for enhanced visualization of the multitemporal image data, facilitating the assessment of the spatiotemporal appearance patterns of lesions. Pixel-mapping functions are used to map individual time-series signals to pseudo-colors which reflect dedicated features of the temporal signal. These signal features are derived from explicit mathematical models of the time-series signals, like in the three-time-points (3TP) method [3] illustrated in Fig. 2.

Figure 2: Model-based 3TP method [3] as the basis for pseudo-coloring of lesions voxels: the intensity of the pseudo-color represents the amount of signal uptake between the precontrast and the early postcontrast image. The presence or absence of a washout is associated with the color hue, yielding thus a simple lesions’ evaluation method which is capable to integrate information from exactly one precontrast and two postcontrast images.

Definition of ROI

The region of interest (ROI) is mostly determined by seeded region growing and also by active contour models that approximate correctly the shapes of organ boundaries. Fig. 3 visualizes such an ROI in breast MRI [1].

Figure 3: Example slices showing the enhancement of lesion tissue over time. The upper left image shows a typical field-of-view as it is used for simultaneous imaging of both breasts. The position of a region-of-interest (ROI) in the right breast is depicted by a white box. The white circle indicates the position of the lesion, which segmentation is presented in white. The remaining images show a magnified view of the ROI in the precontrast (upper right), first postcontrast (lower left) and fifth postcontrast (lower right) image. The lesion exposes a heterogeneous enhancement pattern: the intensity of the tissue in the lesion center continuously increases over time, whereas tissue at the lesion border reaches its peak intensity in the early postcontrast image and exposes lower intensity values in the late postcontrast image. These subtle differences in the temporal characteristics of the tissue are difficult to recognize in the original images but nevertheless important for the differential diagnosis of tumors [1].

Experiments: Classification Results

Two different approaches for accentuating the spatio-temporal appearance pattern of lesions in DCEMRI studies of breasts are presented: unsupervised vector quantization and support vector machines. Fig. 4 shows visualizations based on the outcome of the unsupervised segmentation for a malignant lesion [1]. The images show the cluster distribution for the given slices and at the same time the corresponding time-series signal prototype for each cluster. Thus, a very accurate representation is obtained revealing the nuances in tissue transition. Fig. 5 visualizes the classification performance of the support vector machine compared with a reference label derived from 3TP [1]. All voxels marked by the radiologist exhibit a significant signal enhancement and have to be regarded as suspicious lesion voxels. The marked voxels can be further subclassified into three tissue classes according to the color hue of the pseudo-color assigned by 3TP: red, blue and green voxels indicate malignant, benign and suspicious (with indistinct washout characteristics) time-series signals, respectively.

Both techniques lead to visualizations of temporal image sequences in which the pseudocolor of voxels reflects the temporal characteristic of the underlying tissue. Therewith, subregions with different enhancement kinetics can be identified by means of a single 3D color image, exposing important information about the lesion architecture by the topological pattern of different tissue types in the heterogenous lesion tissue. Unlike the model-based 3TP method which only permits to consider data from one precontrast and two postcontrast images, the proposed techniques are capable of exploiting the information of the entire time-series signals. In consideration of the fact that a wide range of imaging protocols with different spatial and temporal resolutions are used in clinical practice, this flexibility with respect to the input data is a beneficial feature of the adaptive approaches.

Figure 4: Segmentation method applied to multilocullar recurrent ductal carcinoma with four clusters. The left image shows the cluster distribution for each slice ranging from 13 to 16. The right image presents for each cluster the representative time-series signal (displayed as relative enhancement curves) and the corresponding quantities initial signal change (sai) and postinitial signal change (svp). The pseudo-color images indicate a lesion core of malignant tissue exposing signals with strong uptake and distinct washout characteristics [1].
Figure 5: Same tumor as in Figure 4. Four adjacent image slices showing the malignant lesion with pseudo-colors reflecting the local probability of malignant, benign and normal tissue (left 2 x 2 image matrix) or the local tissue classification as malignant (red), normal (green) or benign (blue) tissue (middle 2 x 2 image matrix). The 3TP-based visualization of the lesion is presented in the right 2 x 2 image matrix. All types of pseudocolor visualizations expose the heterogenous structure of the lesion tissue, with malignant tissue areas at the lesion margin and tissue with benign signal characteristics in the core and right part of the lesion [1].

References and Resources

  1. ↑ 1.0 1.1 1.2 1.3 1.4 1.5 1.6 T. Twellmann, A. Meyer-Baese, O. Lange, S. Foo and T. Nattkemper, “Model-free visualization of suspicious lesions in breast MRI based on supervised and unsupervised learning”, Engineering Applications of Artificial Intelligence, 21:129-140, 2008.
  2. ↑ 2.0 2.1 C. K. Kuhl, P. Mielcareck, S. Klaschik, C. Leutner, E. Wardelmann, J. Gieseke, and H. Schild, “Dynamic breast mr imaging: Are signal intensity time course data useful for differential diagnosis of enhancing lesions?”, Radiology, 211:101-110, 1999.
  3. ↑ 3.0 3.1 F. Kelcz, E. Furman-Haran, D. Grobgeld and H. Degani, “Clinical testing of high-spatial-resolution parametric contrast-enhanced MR imaging of the breast,” American Journal of Roentgenology, 179:1485-1492, 2002.

Project

Describe mathematically the three signals from Figure 2. Randomly produce signals within a 5% variation boundary for these three signal categories and fill in three different shapes such as a circle, a star-like structure and a line segment.

Choose a simple unsupervised clustering algorithm (Kohonen map) and determine the segmentation results according to the given signals. Visualize the pseudocolor structure for each shape.

Wiki Assessment

Describe the multitemporal nature of many medical images?

What is an ROI?

What are the differences between time-series classification and traditional stationary features?

Describe how both supervised and unsupervised methods can be applied to classification?

How can be pseudocolors be employed for visualization purposes?

What is a spatiotemporal pattern in an image?

Retrieved from "https://ml.cecs.ucf.edu/meli/wiki/index.php?title=Visualization_of_Lesions_in_Breast_MRI&oldid=545"