Prediction of NSCLS from CT scan and habitual data of a person after two years by using CNN and LSTM machine learning techniques
Principal Investigator
Name
Mayur Munshi
Degrees
M.Sc.
Institution
University of Manchester
Position Title
Part time doctoral student
Email
About this CDAS Project
Study
NLST
(Learn more about this study)
Project ID
NLST-481
Initial CDAS Request Approval
Feb 14, 2019
Title
Prediction of NSCLS from CT scan and habitual data of a person after two years by using CNN and LSTM machine learning techniques
Summary
As a participant to Radiotherapy Machine learning network,I with my team of three machine learning experts intend to use NLST data base of 100 patients diagnosed with NSCLS of various stages.
We would like to use a special kind of recurrent neural networks (RNNs) called Long Short term memory networks (LSTM) that are capable of learning long-term dependencies. LSTMs were introduced by Hochreiter & Schmidhuber(1997).We would like to use demographic and habitual data available in text format to merge with the imaging data of CT scans to predict the probability of a person getting NSCLS in the next two years.
We are taking an inspiration from the Kaggle competition in 2017 where one project was called Predicting Lung cancer outcome.This project's outcome was prediction of lung cancer after one year.
We also want to experiment with different machine learning approaches and hence would like to use LSTM to combine histo-pathology,behavioral data and habitual data of the persons whose CT scans are done as part of lung trial by NLST.we would also like to use Convolution Neural Network(CNN) that will only use Image analyses and compare accuracy of both the predictions.
We would like to use 100 such individual cases(Patients)with complete information of their demographics,habit information and any genetic sequences available. For the initial study population, we would like to request cases who were diagnosed within two years of a positive or negative screen. We would like at least 20 of those patients to be individuals whose screens were negative and yet subsequently developed non-small cell lung cancer, with a particular focus on squamous cell lung cancer. We would also like 100 individuals who were never diagnosed with lung cancer and had at least 24 months of follow-up time following their last screen. We would like relevant demographic data for all requested participants.
We would like to use a special kind of recurrent neural networks (RNNs) called Long Short term memory networks (LSTM) that are capable of learning long-term dependencies. LSTMs were introduced by Hochreiter & Schmidhuber(1997).We would like to use demographic and habitual data available in text format to merge with the imaging data of CT scans to predict the probability of a person getting NSCLS in the next two years.
We are taking an inspiration from the Kaggle competition in 2017 where one project was called Predicting Lung cancer outcome.This project's outcome was prediction of lung cancer after one year.
We also want to experiment with different machine learning approaches and hence would like to use LSTM to combine histo-pathology,behavioral data and habitual data of the persons whose CT scans are done as part of lung trial by NLST.we would also like to use Convolution Neural Network(CNN) that will only use Image analyses and compare accuracy of both the predictions.
We would like to use 100 such individual cases(Patients)with complete information of their demographics,habit information and any genetic sequences available. For the initial study population, we would like to request cases who were diagnosed within two years of a positive or negative screen. We would like at least 20 of those patients to be individuals whose screens were negative and yet subsequently developed non-small cell lung cancer, with a particular focus on squamous cell lung cancer. We would also like 100 individuals who were never diagnosed with lung cancer and had at least 24 months of follow-up time following their last screen. We would like relevant demographic data for all requested participants.
Aims
The specific aim of this project is to predict based on training data and using unstructured machine learning approach, the probability of a person to have most common form of Lung cancer ,NSCLS.
Our second aim is to compare two different techniques of machine learning, CNN and LSTM and match accuracy of their outcomes with each other and individually with actual available results of the patients used as test data for evaluating our prediction model.
Collaborators
Mr. Mayur Munshi,Principal Clinical Scientist-Medical Physicist,Colchester Hospital,ESNEFT
Dr Yemin Wang,University of Sheffield
Mr. Ye Jiang,University of Sheffield
Dr.Shenyuan Ren,Oxford University