Multiple Instance Learning for Breast Cancer Magnetic Resonance Imaging

Abstract

In this thesis we evaluate the efficacy of multiple instance learning (MIL) as a 'pure' machine learning approach for the diagnosis of breast cancer in magnetic resonance images (MRI). The traditional approach for the diagnosis of breast cancer is based on region-of-interest (ROI) based single instance learning (SIL). In the ROI-based SIL, the classification of benign and malignant lesions depends on the features, which are extracted from segmented ROIs. But, an accurate segmentation of a ROI is a challenging task due to poor signal-to-noise-ratio and faint edges due to partial volume effects. Therefore, variations in the segmentation of ROIs can affect the diagnostic outcome. MIL is a relatively new method in the supervised learning, where each sample is represented as a bag of instances. An image in the context of MIL can be considered as a bag of pixels, tiles, or ROIs, which correspond to instances. Here we apply tile-based MIL, where the diagnosis of breast cancer is based on tile-based features which do not require the segmentation of ROIs. Therefore, tile-based MIL has the potential to provide the classification of breast cancer without segmenting the ROIs. This is the motivation for the main objective of this thesis: to estimate the efficacy of tile-based MIL in the detection and the diagnosis of breast cancer MRI.We initially evaluate the potential of MIL for the detection and the diagnosis of breast cancer in anatomical T2-weighted MRI. In particular, we compare the performance of a MIL-based learner, i.e., citation-kNN (CkNN) against conventional kNN and a Random Forest classifier. We utilise both (generic) tile-based spatial features and (domain specific) ROI-based features. We perform experiments on two datasets consisting of 77 mass-like lesions and 129 both mass-like and non-mass-like lesions. The performance of CkNN as both a diagnostic and screening tool is evaluated using the area under the receiver operating characteristic curve (AUC), estimated over 10-fold cross validation. Results demonstrate that the tile-based CkNN has equivalent performance to the ROI-based classification. However, tile-based MIL has an advantage that it does not require the domain specific ROI-based features typically used in breast MRI. This not only has the potential to make the tile-based classification robust to inaccuracies in the delineation of suspicious lesions, but also makes it suitable for the detection of suspicious lesions prior to segmentation.Next, we investigate the performance of CkNN for the diagnosis of breast cancer using dynamic MRI. Specifically we use generic tile-based spatio-temporal features derived from T2-weighted MRI and T1-weighted dynamic contrast enhanced MRI. We utilise a discrete cosine transform and contrast enhancement models as feature extraction techniques. We compare the III performance of CkNN and kNN against a traditional approach based on bespoke ROI-based features using the 77 mass-like lesions. Empirical results show the equivalent classification performance of both ti...