Genetic Algorithms in CAD Mammography

Several research groups have developed computer-aided diagnosis (CAD) programs for the detection and classification of microcalcifications and masses. For most of these programs, there are some common steps that have to be fulfilled in order to find the suspect lesions. Figure 4.1 shows an example of the typical steps needed for a CAD program. Starting from a digital (or digitized) mammogram, the first operations are the preprocessing ones. Here, the breast is segmented and some filtering or normalization accomplished in order to improve the quality of the image and reduce the noise. Then, a signal extraction step is performed. In this phase, objects similar to the lesions are isolated by means of different techniques. After that, a set of features is calculated on the extracted signals. Basically, researchers have investigated two types of features: those traditionally used by radiologists (gradient-based, intensity-based, and geometric features) and high-order features that may not be as intuitive to radiologists (e.g., texture features). Finally, a classification (false-positive reduction) step is performed, where on the basis of the mentioned features, false signals are separated from the suspect lesions by means of a classifier. In other words, the candidate lesions are first located and then further analyzed in a feature analysis and classification phase to determine the final classification of each candidate. Each stage of most of the CAD schemes uses multiple parameters such as threshold values, filter weights, and region of interest (ROI) sizes. To have a high performance, the values of these parameters need to be selected optimally. In general, however, the optimal set of parameters may change when a component of the imaging chain is modified or changed. This is because some of the parameter values depend implicitly or explicitly on the previous steps. Also, the parameter values have to be redetermined if a new component is added to the CAD scheme for improvement of its performance. Many CAD systems are composed of several independent yet interrelated parts, and some optimization studies have to be done for maximizing the performance. A commonly used approach is to try different combinations of parameters in an ad hoc manner and empirically select the best values based on the test results. However, this manual optimization process searches only very limited regions of the large-dimensional parameter space. In order to overcome the difficulties associated with manual optimization, automated methods have been developed.