Future diagnosis prediction of lung cancer using convolutional and recurrent networks.
Principal Investigator
Name
Latifur Khan
Degrees
M.S, Ph.D
Institution
University of Texas at Dallas
Position Title
Professor
Email
About this CDAS Project
Study
NLST
(Learn more about this study)
Project ID
NLST-449
Initial CDAS Request Approval
Oct 31, 2018
Title
Future diagnosis prediction of lung cancer using convolutional and recurrent networks.
Summary
Problem Statement: Given a sequence of chest CT scans for a patient with annotations from radiologists, we predict the next possible annotation in this sequence.
Motivation: In 2017, there will be an estimated 1,688,780 new cancer cases diagnosed and 600,920 cancer deaths in the US. We believe that solving the problem statement mentioned above, will give an idea to the doctors whether the current treatment to a patient is helping cure lung cancer and, early diagnosis of lung cancer in a patient.
Proposed method: Use 3 convolutional networks to extract features from the CT scans and use a recurrent network to understand the sequence between the three extracted features which differ by time.
Details: We would train a network to learn features from CT scan images of a patient taken at different years (or in general time), to map to diameter of non-calcified nodules / masses using convolutional neural networks and use multiple convolutional neural networks as units in recurrent neural network to predict the diameter of the non-calcified nodules / masses to be seen in the next CT scan. This would enable doctors to better diagnosis patients undergoing treatment for cancer to see if the treatment is beneficial to the patient. And, with knowing the non-calcified nodules / masses diameter of a patient in the future scans would help in early diagnosis of cancer and enables doctors to make early measures for these patients. It would be of great value in detection and treatment of cancer if we are able to predict what information the next CT scan would hold after some amount of time (probably a year).
Future Work: To predict the time when a patient may be diagnosed with lung cancer and, also to predict after certain time from now, what information their CT scan would hold.
Motivation: In 2017, there will be an estimated 1,688,780 new cancer cases diagnosed and 600,920 cancer deaths in the US. We believe that solving the problem statement mentioned above, will give an idea to the doctors whether the current treatment to a patient is helping cure lung cancer and, early diagnosis of lung cancer in a patient.
Proposed method: Use 3 convolutional networks to extract features from the CT scans and use a recurrent network to understand the sequence between the three extracted features which differ by time.
Details: We would train a network to learn features from CT scan images of a patient taken at different years (or in general time), to map to diameter of non-calcified nodules / masses using convolutional neural networks and use multiple convolutional neural networks as units in recurrent neural network to predict the diameter of the non-calcified nodules / masses to be seen in the next CT scan. This would enable doctors to better diagnosis patients undergoing treatment for cancer to see if the treatment is beneficial to the patient. And, with knowing the non-calcified nodules / masses diameter of a patient in the future scans would help in early diagnosis of cancer and enables doctors to make early measures for these patients. It would be of great value in detection and treatment of cancer if we are able to predict what information the next CT scan would hold after some amount of time (probably a year).
Future Work: To predict the time when a patient may be diagnosed with lung cancer and, also to predict after certain time from now, what information their CT scan would hold.
Aims
- Predict the information such as non-calcified nodules / masses to be obtained in a future CT scan of patient based on previous CT scans.
- With this information, doctors will have an idea how the current treatment for a cancer patient will react on that patient.
- People who has history of smoking, or smokers, or history of cancer are probable to get cancer in the future. With the obtained future CT scan information, we can diagnose whether a person will have cancer earlier.
Collaborators
Pillai, Arvind - atp170130@utdallas.edu
Gopalakrishnan Sakunthala, Mukesh Kumar - mxg165930@utdallas.edu
Dr. Latifur Khan - lkhan@utdallas.edu