a) Perform pixel labeling for the lung area. The labeled lung area forms a 3D region in a chest CT scan. Subsequent processing and analysis steps will be restricted to the labeled lung area and its neighborhood. We plan to apply state-of-the-art deep learning based semantic image segmentation algorithms to pixel labeling for the lung area.

b) Detect nodules in the lung area. Each nodule forms a local region. The pixels inside each nodule are labeled and the bounding box of every nodule is also calculated. We would like to obtain as many CT scans with lung nodules as possible for achieving this goal because deep neural networks require a large amount of training data.

For lung nodule detection, we plan to apply the latest deep learning based object detection methods. We also plan to adapt a state-of-the-art salient object detection network developed by the PI to lung nodule detection due to the close connection between these two detection problems.

c) Determine whether lung cancer has developed or will develop by observing the detected nodules. When there are insufficient evidences to conclude that a cancer has developed, predicts whether a cancer will develop within 6-12 months. We would like to obtain all CT scans that were diagnosed with lung cancer as well as additional CT scans from cancer free patients.

We plan to apply the latest deep learning based image classification methods, including the Residual Network. We also plan to adapt a state-of-the-art image description method developed by the PI to further improve the classification performance.

d) Characterize benign and malignant lung nodules on the basis of information extracted from CT scans as well as other patient-specific clinical records. The aforementioned image classification and description methods can also be applied to local image regions containing nodules.

none.