Robust Radiomics Feature Extraction for Lung Cancer
Principal Investigator
Name
Suthirth Vaidya
Degrees
M.Tech, B.Tech
Institution
Predible Health
Position Title
CEO
Email
About this CDAS Project
Study
NLST
(Learn more about this study)
Project ID
NLST-318
Initial CDAS Request Approval
Jun 20, 2017
Title
Robust Radiomics Feature Extraction for Lung Cancer
Summary
Radiomics, the process of extracting high-throughput mineable data from digital medical images to reveal insights about underlying pathophysiology has been rapidly gaining deep interest from the clinical and research community. Various publications have established the correlation of radiomic features with the prognosis and mutation status in Lung cancer. In this fast-paced research field, a common issue faced by several groups is the reproducibility of the generated features, and hence the established correlations.
Radiomic features are extracted from the lung nodule and pathological regions, using carefully delineated segmentations. Due to vague and intricate boundaries, humans often face difficulty in the clear delineation of the margins. This has led to increased exploration of ‘semantic features’ which do not require segmentations for analysis. However, the quantitative nature of ‘agnostic features’, which are extracted from the segmentations can lead to better modeling of the required outcomes.
It is in this context that the relevance of Deep Learning and its potential in robust segmentation from medical images can be explored. Using methods which we have earlier published for the robust segmentation of Multiple Sclerosis lesions and Brain Tumors, we propose to build 3D Convolutional Neural Network (CNN) models which can robustly segment nodules and pathological regions in the lung, without any human intervention. The segmentations form the basis for high-throughput radiomic feature extraction.
This study then explores to use the robust CNN-based Radiomic feature extractors to establish the correlation of Lung CT images at various stages with their progression and outcomes.
Radiomic features are extracted from the lung nodule and pathological regions, using carefully delineated segmentations. Due to vague and intricate boundaries, humans often face difficulty in the clear delineation of the margins. This has led to increased exploration of ‘semantic features’ which do not require segmentations for analysis. However, the quantitative nature of ‘agnostic features’, which are extracted from the segmentations can lead to better modeling of the required outcomes.
It is in this context that the relevance of Deep Learning and its potential in robust segmentation from medical images can be explored. Using methods which we have earlier published for the robust segmentation of Multiple Sclerosis lesions and Brain Tumors, we propose to build 3D Convolutional Neural Network (CNN) models which can robustly segment nodules and pathological regions in the lung, without any human intervention. The segmentations form the basis for high-throughput radiomic feature extraction.
This study then explores to use the robust CNN-based Radiomic feature extractors to establish the correlation of Lung CT images at various stages with their progression and outcomes.
Aims
1: Study and establish the various co-morbidities or abnormalities commonly present in Lung Cancer patients of different stages
2: Develop segmentation models for nodules and other commonly present abnormalities using Convolutional Neural Networks from the training dataset
3: Validate the performance of the segmentation models using an independent validation dataset. This will ensure the robustness of the feature extractors
4: Using the radiomics features, study correlations with outcomes and disease progression over time
Collaborators
None outside of Predible Health
Related Publications
-
Unboxing AI - Radiological Insights Into a Deep Neural Network for Lung Nodule Characterization.
Venugopal VK, Vaidhya K, Murugavel M, Chunduru A, Mahajan V, Vaidya S, Mahra D, Rangasai A, Mahajan H
Acad Radiol. 2020 Jan; Volume 27 (Issue 1): Pages 88-95 PUBMED