Aria Pezeshk

PhD

FDA

Research Fellow

NLST (Learn more about this study)

NLST-109

Jan 13, 2015

Deep learning and compact feature representations for medical image analysis

Deep learning is a recent approach in machine learning that is well-suited for complex classification tasks which require highly nonlinear behavior and large learning capacity. Unlike traditional techniques, deep learning is highly scalable to big data problems which make it the ideal tool for scouring the enormous quantity of medical data collected every day. Artificial intelligence methods traditionally extract useful attributes of the data using feature extraction techniques. In addition, to learn data patterns, these methods require data labeling, e.g., training cases that are labeled as diseased versus non-diseased. In contrast, deep learning techniques can utilize raw unlabeled data to automatically discover critical patterns for classification and can therefore significantly reduce the burden in preparation and pre-processing of data by clinical experts. Since its introduction only a few years ago, this technique has outperformed other approaches in artificial intelligence by wide margins across a large range of applications, and reached human levels of accuracy for the first time in several tasks. These remarkable results have led to a rapid expansion of deep learning into both commercial and research settings. Only a few studies have so far explored using this technique for medical applications. In this project we will focus on using deep learning across several medical application areas in order to fill the knowledge gap and to demonstrate the multi-functional capabilities of this promising technology in solving problems that can benefit public health both in common practice of medicine and for development of medical countermeasures during emergencies. In our first study in the series, we will use two deep learning architectures referred to as restricted Boltzmann machines (RBMs), and deep convolutional neural networks (CNNs) to investigate learning 3D features from large libraries of chest CT scans in a computer aided diagnosis (CAD) system for pulmonary nodule detection. We plan to pre-train the deep learning models using the large number of scans available from the NLST dataset, and subsequently test the performance of a CAD system trained on the learned features on labeled cases from the LIDC dataset.
In a second study, we plan to build a deep-learning system for characterization of dataset complexity. An important issue in comparing the performance of different CAD systems is that different systems are often tested on distinct datasets that contain different levels of case or lesion complexity. If lesions in different datasets can be grouped into categories such as difficult, medium and easy to detect, comparison of two systems within each category can be more meaningful. Deep learning is suited for this categorization because it can discover useful features given unlabeled data, i.e., sets of lesions without a complexity rating, followed by supervised training with a small set of lesions with complexity ratings from a panel of experts. In our second study, we will use the learned 3D feature to train a system that can accurately predict a subtlety rating for lesions in a dataset. The predicted ratings from different datasets can then be compared to assess their corresponding levels of complexity.