Lung nodules identification in CT scans using multiple instance learning.

Abstract

Computer Aided Diagnosis (CAD) systems for lung nodules diagnosis aim to classify nodules into benign or malignant based on images obtained from diverse imaging modalities such as Computer Tomography (CT). Automated CAD systems are important in medical domain applications as they assist radiologists in the time-consuming and labor-intensive diagnosis process. However, most available methods require a large collection of nodules that are segmented and annotated by radiologists. This process is labor-intensive and hard to scale to very large datasets. More recently, some CAD systems that are based on deep learning have emerged. These algorithms do not require the nodules to be segmented, and radiologists need to only provide the center of mass of each nodule. The training image patches are then extracted from volumes of fixed-sized centered at the provided nodule's center. However, since the size of nodules can vary significantly, one fixed size volume may not represent all nodules effectively. This thesis proposes a Multiple Instance Learning (MIL) approach to address the above limitations. In MIL, each nodule is represented by a nested sequence of volumes centered at the identified center of the nodule. We extract one feature vector from each volume. The set of features for each nodule are combined and represented by a bag. Next, we investigate and adapt some existing algorithms and develop new ones for this application. We start by applying benchmark MIL algorithms to traditional Gray Level Co-occurrence Matrix (GLCM) engineered features. Then, we design and train simple Convolutional Neural Networks (CNNs) to learn and extract features that characterize lung nodules. These extracted features are then fed to a benchmark MIL algorithm to learn a classification model. Finally, we develop new algorithms (MIL-CNN) that combine feature learning and multiple instance classification in a single network. These algorithms generalize the CNN architecture to multiple instance data. We design and report the results of three experiments applied on both generative (GLCM) and learned (CNN) features using two datasets (The Lung Image Database Consortium and Image Database Resource Initiative (LIDC-IDRI) \cite{armato2011lung} and the National Lung Screening Trial (NLST) \cite{national2011reduced}). Two of these experiments perform five-fold cross-validations on the same dataset (NLST or LIDC). The third experiment trains the algorithms on one collection (NLST dataset) and tests it on the other (LIDC dataset). We designed our experiments to compare the different features, compare MIL versus Single Instance Learning (SIL) where a single feature vector represents a nodule, and