Thermography-based Breast Abnormality Detection using Siamese Network

Abstract

Thermography is a potential imaging modality for early breast abnormality detection. Both breasts of a healthy individual have a similar temperature distribution. Thermal asymmetry between the left and the right breasts may indicate the presence of an abnormality. This work introduces a novel Siamese network for learning the similarity between the left and the right breast thermogram images to identify breast abnormalities. This work also proposes a novel algorithm to identify breast abnormalities using the similarity scores obtained from the introduced Siamese network. A comparison of the proposed breast abnormality detection methodology at different margin values is presented. The proposed methodology using the Siamese network achieves an accuracy of 81% with a standard error of 0.3% when the margin was set to 1. All evaluations are done using a publicly available dataset.