Final Considerations

To reduce the complexity of the proposed algorithm, a multiresolution scheme was used. Thus, the initial search is done on a subsampled image and only the potential points found in this resolution (points where $\Lambda(\psi^{s,\xi,d},Y)$ is greater than a threshold) are used in subsequent steps. These consist of incrementing the resolution of the image and, at these potential points $(p_x,p_y)$ , re-calculate the parameters of the template (size and $\xi$ ) by using a gradient descent algorithm. With this new template a refinement of these potential points is done. This loop is repeated until the original size of the image is reached or there are no potential points in the image (no mass in the mammogram).

When the algorithm has finished, a set of regions of the mammogram are marked as suspicious RoIs. Each RoI consists of a centre and a surrounding box indicating the size of the possible mass. Figure  shows the centre of the suspicious RoIs found in different mammograms. We can recover also the final shape of the suspicious RoI. However, as the training step is done using rough annotations, the final shape of the template provides poor information.

The performance of the algorithm is qualitatively shown in Figure . Each square centre corresponds to the centre of the found mass, while the square size represents the mass bounding square. Note that a large number of the detected RoIs actually correspond to normal tissue. Thus, a subsequent step will be necessary in order to reduce the number of false positives.

Figure 4.7: Centre and size of the suspicious regions found in four mammograms of the MIAS database [169] with clear masses (marked in white).

(a)	(b)	(c)	(d)