Multiple Instance Learning for Breast Cancer Magnetic Resonance Imaging

In this paper we evaluate the suitability of multiple instance learning (MIL) for the classification of T2 weighted magnetic resonance images (MRI) of the breast. Specifically, we compare the performance of citation-kNN against traditional kNN and a random forest (RF) classifier. We utilise both (generic) tile-based features and (domain specific) region-of-interest (ROI) based features We perform experiments on two datasets consisting of A) mass-like lesions and B) both mass-like and non-mass-like lesions. The performance of citation-kNN as both a diagnostic and screening tool is evaluated using the area under the receiver operating characteristics curve (AUC), estimated over 10-fold cross-validation. Results demonstrate that citation- kNN has equivalent performance to traditional kNN and RF. However, the tile-based approach used by citation-kNN does not require the domain specific ROI-based features typically used in breast MRI. This not only makes citation-kNN robust to inaccuracies in the delineation of suspicious lesions, but also makes it suitable for use as a screening tool, where the aim is to discriminate lesions from normal tissue.