Study
NLST
(Learn more about this study)
Project ID
NLST-871
Initial CDAS Request Approval
Jan 26, 2022
Title
Fully automated end to end analysis of non-small-cell lung carcinoma using deep learning techniques
Summary
Lung cancer is one of the critical cancers as it cannot be accurately diagnosed at an early stage leading to most cancer deaths worldwide. The non-small-cell lung carcinoma (NSCLC) is the most common type of lung cancer. The main goal of this project is to develop a fully automated deep learning methodology for lung cancer diagnosis using computed tomography (CT) images for monitoring its growth/regression over time to evaluate treatment efficacy for the cancer in the patients. A U-Net based, segmentation model available will be used as a preprocessing step to segment the lungs from the CT images. The segmented lungs will be further used for tumour identification and segmentation by developing a 3D Deep Convolutional Neural Network (DCNN) based on the state-of-the-art deep learning architectures (e.g. UNet++, Attention U-Net, GAN). The classification of patient’s five-year survival will be performed using a time series analysis based on deep learning network and based on conventional methods using tumour size, tumour growth over time and mean intensity of tumour. The results obtained will be compared with the existing medical to medical transfer learning model for a detailed performance assessment.
Aims
The general objective of this research project is to develop a fully connected 3D CNN to detect the tumor in the NSCLC patients to detect cancerous organ from the normal ones. The tumours will then be segmented from the CT images and the monitoring of the tumour progression/regression will be analysed using longitudinal data across several time points. The prognostication on the tumour will be performed by developing a deep learning-based imaging biomarker. The sub-objective and the activities are as follows:
1.Preprocessing of the data set to be used
2.Performing lung segmentation using the pre-existing model available
3.Developing the 3D state of art deep learning model for tumor detection in NSCLC patients
4.Prediciton Survival using time series based on deep learning and conventional methods
5.Comparing the developed model with a state-of-the-art model like MED-3D (developed based on transfer learning approach)
Collaborators
Jenita Manokaran, PhD Student
