Pre-hoc and Post-hoc Diagnosis and Interpretation of Breast Magnetic Resonance Volumes

Abstract

Breast cancer is among the leading causes of death in women. Aiming at reducing the number of casualties, breast screening programs have been implemented to diagnose asymptomatic cancers due to the correlation of higher survival rates with earlier tumour detection. Although these programs are normally based on mammography, magnetic resonance imaging (MRI) is recommended for patients at high-risk. The interpretation of such MRI volumes is timeconsuming and prone to inter-observer variability, leading to missed cancers and a relatively high number of false positives provoking unnecessary biopsies. Consequently, computeraided diagnosis systems are being designed to help improve the efficiency and the diagnosis outcomes of radiologists in breast screening programs. Traditional automated breast screening systems are based on a two-stage pipeline consisting of the localization of suspicious regions of interest (ROIs) and their classification to perform the diagnosis (i.e. decide about their malignancy). This process is typically ineffective due to the usual expensive inference involved in the exhaustive search for ROIs and the employment of non-optimal hand-crafted features in both stages. These issues have been partially addressed with the introduction of deep learning methods that unfortunately need large strongly annotated training datasets (voxel-wise labelling of each lesion), which tend to be expensive to acquire. Alternatively, the use of weakly labelled datasets (i.e volume-level labels) allows diagnosis to become a supervised classification problem, where a malignancy probability is estimated after examining the entire volume. However, large weakly labelled training sets are still required. Additionally, to facilitate the adoption of such weakly trained systems in clinical practice, it is desirable that they are capable of providing the localization of lesions that justifies the automatically produced diagnosis for the whole volume. Nonetheless, current methods lack the precision required for the problem of weakly supervised lesion detection. Motivated by these limitations, we propose a number of methods that address these deficiencies. First, we propose two strongly supervised deep learning approaches that not only can be trained with relatively small datasets, but are efficient in the localization of suspicious tissue. In particular, we propose: 1) the global minimization of an energy functional containing information from the semantic segmentation produced by a deep learning model for lesion segmentation, and 2) a reinforcement learning model for suspicious region detection. Diagnosis is performed by classifying suspicious regions yielded by the reinforcement learning model. Second, aiming to reduce the burden associated to strongly annotating datasets, we propose a novel training methodology to improve the diagnosis performance on systems trained with weakly labelled datasets that contain a relatively small number of training samples. We further propose a novel 1-class saliency detector to automatically localize lesions associated with the diagnosis outcome of this model. Finally, we present a comparison between both of our proposed approaches for diagnosis and lesion detection. Experiments show that whole volume analysis with weakly labelled datasets achieves better performance for malignancy diagnosis than the strongly supervised methods. However, strongly supervised methods show better accuracy for lesion detection.