Biomedical Signal Analysis: Contemporary Methods and Applications

10. Classification of Dynamic Breast MR Image Data

10 Classification of Dynamic Breast MR Image Data Breast cancer is the most common cancer among women. Magnetic resonance (MR) is an emerging and promising new modality for detection and further evaluation of clinically, mammographically, and sonographically occult cancers [115, 293]. However, film and soft-copy reading and manual evaluation of breast MRI data are still critical, time–consuming and inefficient, leading to a decreased sensitivity [204]. Furthermore, the limited specificity of breast MR imaging continues to be problematic . Two different approaches are mentioned in literature [145] aiming to improve the specificity: (1) single–breast imaging protocols with high spatial resolution offer a meticulous analysis of the lesion’s structure and internal architecture, and are able to distinguish between benign and malignant lesions; (2) lesion differential diagnosis in dynamic protocols is based on the assumption that benign and malignant lesions exhibit different enhancement kinetics. In [145], it was shown that the shape of the time-signal intensity curve is an important criterion in differentiating benign and malignant enhancing lesions in dynamic breast MR imaging. The results indicate that the enhancement kinetics, as shown by the time-signal intensity curves visualized in figure 10.1, differ significantly for benign and malignant enhancing lesions and thus represent a basis for differential diagnosis. In breast cancers, plateau or washout time courses (type II or III) prevail. Steadily progressive signal intensity time courses (type I) are exhibited by benign enhancing lesions. Also, these enhancement kinetics are shared not only by benign tumors but also by fibrocystic changes [145]. Concurrently, computer–aided diagnosis (CAD) systems in conventional X–ray mammography are being developed to expedite diagnostic and screening activities. The success of CAD in conventional X–ray mammography motivated the research of similar automated diagnosis techniques in breast MRI. Although, they are an issue of enormous clinical importance with obvious implications for health care politics, research initiatives in this field concentrate only on pattern recognition methods based on traditional artificial neural networks [161] ,[1, 162, 271]. A standard multilayer perceptron (MLP) was applied to the classi- fication of signal–time curves from dynamic breast MRI in [161]. The 276 Chapter 10 t signal intensity [%] Ib II III Ia early intermediate and late postcontrast phase Figure 10.1 Schematic drawing of the time-signal intensity curve types [145]. 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 SIc−SI SI where SI is the precontrast signal intensity and SIc is the postcontrast signal intensity. In breast cancers, plateau or washout time courses (type II or III) prevail. Steadily progressive signal intensity time courses (type I) are exhibited by benign enhancing lesions. major disadvantage of the MLP approach and also of any other supervised technique is the fixed number of input nodes, which imposes the constraint of a fixed imaging protocol. Delayed administration of the contrast agent or a different temporal resolution has a negative effect on the classification and segmentation capabilities. Thus, a change in the MR imaging protocol requires a new training of the CAD system. In addition , the system fails in most cases to diagnose small breast masses with a diameter of only a few millimeters. It must be mentioned that during the training phase of a classifier, a histopathologically classified lesion represents only a single input pattern. There is an urgent need, based on the limited number of existing training data, to efficiently extract information from a mostly inhomogeneous available data pool. While supervised classification techniques often fail to accomplish this task, the proposed biomimetic neural networks, in the long run, represent the best training approaches leading to advanced CAD systems. When applied to segmentation of MR images, traditional pattern recognition techniques such as the MLP have shown unsatisfactory detection results and limited application capabilities [1, 162]. Furthermore, the underlying supervised nonbiological learning strategy leads to the inability to capture the feature structure of the breast lesion in the neural architecture. One recent paper demonstrated examples of the segmentation of dynamic breast MRI data sets by unsupervised neural networks. Classification of Dynamic Breast MR Image Data 277 Trough use of a Kohonen neural network, areas with similar signal time courses in...