Seungho Lee

M.S.

Vuno

Researcher

NLST (Learn more about this study)

NLST-322

Jul 10, 2017

Enhancement of pulmonary nodule detection algorithm using large scale LDCT data

Early detection of pulmonary nodule is thought to be one of the most effective ways to reduce mortality from lung cancer. Especially, the use of low-dose computed tomography (LDCT) has proven effective in reducing mortality from lung cancer. As a result, lung cancer screening program for high-risk subjects are steadily expanding, and it is obvious that a huge amount of LDCT scans will be provided to medical field. At this point, Automated process of detecting pulmonary nodules can contribute to increasing efficiency and cost savings of treats of lung cancer using LDCT. With the rapid development of modern computer vision and machine learning, we expect to subject better solutions for this problem by using the NLST datasets.

In this project, we are trying to develop a method that would improve the existing nodule detection algorithm. In particular, We focus on the difficulties which rise in obtaining Ground Truth data that describes the location of the pulmonary nodules from 3D dataset as CT scans. By using modern computer vision method and deep learning approach, we would build algorithms that extracts useful features from CT scans which is not fully annotated for lesions location. With this feature, we also may build methodologies to enhance the existing nodule detection algorithm. The success of this project would provide a new insight into making use of abundantly available unlabeled LDCT scans, as well as increasing the efficiency of lung cancer screening.

Aim 1 : Establish clustering and feature extraction method of LDCT scans which is not fully annotated for lesions location using deep learning technique. Furthermore, focus on developing more efficient methods to combine image data with various serial data to create richer features. We also plan to generate our own annotation for learning the algorithm.

Aim 2 : Apply modern computer vision and deep learning technique to develop an enhancing method to improve nodule detection algorithm trained through supervised learning. This can be achieved by associating the dataset constructed in aim 1 with fully annotated dataset.