Structured Micro-Pattern Based LBP Features for Classification of Masses in Dense Breasts

Cancerous masses detection in dense background is a particularly challenging task for even experienced radiologists due to their similarity of intensity with the overlapped normal dense tissues, obscured boundaries and low contrast between mass and surrounding regions. This paper proposes a novel approach for the identification of cancerous regions located in a dense part of a breast. Careful analysis of nine structured micro-patterns generated using LBP technique revealed the most prominent ones, allowing for successful classification of cancerous regions. The proposed approach was evaluated using two mammographic databases: the publicly available Digital Database for Screening Mammography (DDSM), and a local database of mammograms. A total of 535 Regions of Interest (ROIs) were used (301 ROIs extracted from DDSM and 234 from local database). All 535 ROIs were localized in dense backgrounds of breasts. The experiments showed that features generated from structured micro-patterns can produce very effective and efficient texture descriptors of cancerous ROIs. With only 4 features we obtained an AUC score of 0.957 for DDSM and 0.891 for local dataset using Fischer Linear Discriminant Analysis (LDA) classifier.