Eigenmasses and Eigenrois

We can establish a significant parallelism between face images of the same person and mammographic RoIs images. Two of the most common problems in face recognition are related to illumination and pose changes. Note that, in mammography, we can also talk about variations in illumination and pose. Thus, changes in illumination are related to the acquisition parameters (number and energy of X-rays that go through the breast, the exposure time, the film sensitivity, etc) as well as to the internal density of the breast. On the other hand, changes in the pose can be explained twofold as changes in the global mammogram or in the RoI. Changes of pose in the mammogram are related to the different compression suffered by the breast when the mammogram is acquired. Thus, the shape of a mass, as well as the shape of other internal structures, can be different according to the degree of such compression. Looking at a RoI level, changes in pose can be seen as changes of size and shape of the masses.

Although this parallelism, the transition from face recognition to mass detection is far from trivial, due to the explained previous changes. Namely, the main drawbacks of applying the eigenfaces approach to the detection of masses are the variance of the grey-level range and the variable size of the RoIs. Note that the size of the RoIs depends on the size of the (possible) mass, and there is a huge range of mass sizes [94].

Grey-level and texture variation of RoIs are mainly related to the variation of the acquisition parameters (exposure time, X-ray energy) of mammograms obtained at different time intervals and also to the nature of the breast (breast density and thickness). Using a commonly used simplification, these parameters are considered to affect only to the range of the grey-level values of each RoI. Thus, a solution to take these variations into account can be easily computed by equalizing the images. In this sense, we assume a uniform distribution model. On the other hand, and in contrast with face recognition where a database of faces of the same size is available, the size of the RoIs is not always the same. In order to deal with RoIs of variable size, different proposals can be considered:

1. Cropping the RoI by a size determined by the largest mass in the database.
2. To crop according to the size of the mass, and resize it to obtain equal RoI size.
3. To cluster the database of RoIs in different groups according to their size. Thus, when a new RoI is detected, depending on their size, the corresponding cluster will be used.

Experiments have shown that the best results are obtained using the third approach [130]. Note that the main drawback of this approach is the need of a classification of each RoI into the different RoI size clusters. Figure  shows originals and eigenmasses found using a cluster of RoIs. Note that the eigenmass images show a central mass with brighter grey-level than the surrounding tissue.

To avoid confusion, we shall distinguish between eigenrois and eigenmasses. If all RoIs of the training database contain a mass, the result of the algorithm will be called eigenmasses, while if the database is composed by RoIs with masses and RoIs with normal tissue, we will call the resulting images eigenrois. We will discuss the eigenrois in Chapter .

Figure 4.3: The upper row shows four RoIs corresponding to manually detected masses of similar size, while the lower row shows the corresponding obtained eigenmasses.