Scattering transform for breast cancer detection

López Tenza, Aina

Scattering transform for breast cancer detection

Llicència

Citació

López Tenza, A. Scattering transform for breast cancer detection. 2015. handle: http://hdl.handle.net/10230/25405

Aquesta citació s'ha generat automàticament.

Resum

This project investigates the feasibility of using the scattering transform for breast cancer detection from mammograms. The Scattering Transform is a recently introduced descriptor that follows the scheme of a neural network, whose filters are wavelets followed by a modulus operator. We built a classification scheme that involves a feature extraction and a classification stages. For the first stage, we have used the Scattering Transform for extracting discriminant features robust to small deformations in the image. Moreover, we have also used another descriptor as a reference, the Scale-Invariant Feature Transform, widely used in classification problem. For the second stage, we have trained two models, a generative (based on affine spaces) and a discriminative (based on support vector machines). We tested our pattern recognition system on 1590 mammograms (964 normal cases,/n425 cancer cases and 201 benign cases), achieving more than 90% of accuracy. These results confirm that the Scattering Transform is a powerful tool to describe mammograms.