Kayhan Batmanghelich

PhD

University of Pittsburgh

Assistant Professor

NLST (Learn more about this study)

NLST-363

Nov 7, 2017

Unsupervised Deep Learning Approach to Characterize the Normal-Appearing Tissue in Lung CT images

Deep learning has resulted in significant leaps in the performances of many tasks in medical imaging and computer vision, such as image segmentation and classification. Recently, unsupervised deep learning has produced the state-of-art results for the task of probability density estimation. A robust probability density estimator characterizes complex patterns in data efficiently. In our application, we use a robust probability density estimator to describe the patterns of the healthy appearing tissue in CT images. The success of such approach, which is based on deep learning, heavily depends on the availability of a massive scale dataset such as NLST CT images. Upon the achievement of our aims, we can distinguish abnormal-appearing tissue by comparing its image pattern with the patterns of normal-appearing tissue. NLST is unique dataset since it provides structured annotations, demographic and outcomes. Our goal is to develop a method that can be used for lung screening.

We first develop a supervised deep learning approach to predict the annotation provided by the NLST dataset. As a first task, we use NLST and a few other public datasets to train a supervised deep learning model for nodule detection. Then, we used the trained supervised model as an initialization for the unsupervised model. Finally, we evaluate our method with other clinical measures provided by the NLST dataset.

Aim 1 (Supervised Approach): We develop a supervised approach to predicting the clinical annotations provided in the NLST dataset.

Aim 2 (Unsupervisedn Approach): We train our unsupervised density estimator on the tissue without abnormal annotation. We will evaluate our method on how well it can predict the areas of lung with abnormal annotation.

Aim 3 (Evaluation): We evaluate our unsupervised model by predicting clinical variables provided in the NLST dataset.