Convolutional Siamese Networks for Myocardial Lesion Classification in T1 Maps

This paper presents a novel framework for myocardial lesion classification in T1 maps. Late gadolinium enhancement (LGE) imaging is the gold standard for detecting myocardial lesions, however, it requires contrast injection, which raises two main challenges: i) increased scan times and patient preparation, and ii) acute side effects in a certain class of patients. It would therefore be desirable to use a less invasive method, such as T1 mapping, however, this modality does not always present noticeably increased T1 values in the lesion. Taking into account the above challenges, we propose a two stage framework: i) the approach is able to learn associations between T1 map and LGE image during a training phase; (ii) in a test phase, only the T1 map is used, using the previously learned associations. The associations are learned following three siamese inference methodologies. Our experimental results testify the usefulness of the proposed approach on classification of the myocardium lesion in T1 maps.