Meritxell Tortajada, Presenter: Nothing to Disclose

Arnau Oliver PhD, Abstract Co-Author: Nothing to Disclose

Yago Díez, Abstract Co-Author: Nothing to Disclose

Robert Marti PhD, Abstract Co-Author: Nothing to Disclose

Joan C. Vilanova MD, Abstract Co-Author: Nothing to Disclose

Jordi Freixenet PhD, Abstract Co-Author: Nothing to Disclose

To analyze the performance of a previous developed mass detection single-image CAD system when including information from the comparison of left and right mammograms (dual-image CAD system).

In a previous work we presented a model-based single-image CAD system for mass detection consisting in two different steps (Freixenet et al. Med Phys '09). First of all, the system learns the morphology from a training set of real masses and uses this information to create a set of mass templates and their possible deformations. In the second step, these templates and deformations are used in a probabilistic matching scheme providing a set of suspicious regions. The dual-image CAD system is developed including information from the automatic registration of contralateral mammographic images. An intensity-based non-rigid registration approach which uses mutual information as the optimization method is applied. The comparison between left and right mammograms is integrated into the probabilistic CAD detection framework as prior information. The difference between the registered images is used to enhance the image intensities. Therefore, the structures of the analyzed mammogram that appear brighter than in the contralateral image are highlighted, while the structures that have a similar or darker appearance are obscured. To evaluate the performance of the dual-image CAD, a subset of 120 digitized MLO mammograms and their corresponding contralateral image of the same view were extracted from the MIAS database. A total of 26 of these mammograms contained masses, while the rest were normal ones. The quantitative evaluation of our experiments was done in terms of FROC analysis.

The sensitivities for the single-image CAD system were 80% and 88% at 1.68 and 1.85 false positives per image, respectively. For the dual-image CAD system and at the same sensitivities the false positives per image were 0.90 and 0.99, respectively. The obtained results show the benefits of including the bilateral information.

We successfully integrate bilateral information into our single-image CAD system. The experiments show that the dual-image CAD system reduces the number of false positives per image.

We show that the performance of a CAD system for mammographic mass detection significantly improves by including bilateral comparison (left and right breasts) after image registration.

Tortajada, M, Oliver, A, Díez, Y, Marti, R, Vilanova, J, Freixenet, J, Integrating Bilateral Information in the Eigendetection CAD Approach.  Radiological Society of North America 2010 Scientific Assembly and Annual Meeting, November 28 - December 3, 2010 ,Chicago IL. http://archive.rsna.org/2010/9010702.html