Quantitative radiomic assessment of lung nodules with low dose CT
This project aims to develop a fully automated software tool to quantify the nature of the lung nodules identified in LDCT and predict its probability of being malignant or benign using advanced machine learning techniques and statistical methods with combination of radiomic features of lung nodules and lung parenchyma and their temporal changes in serial follow-up exams, patient demographic information, and risk factors from medical, occupational or smoking history.
1. Develop robust methods for lung segmentation, chest wall and rib segmentation, nodule segmentation, and other lung disease segmentation – Although we have developed these segmentation methods in our previous studies, we have not yet evaluated their performances in low-dose CT images. Because of the higher noise and possible lower image quality in the LDCT images, these methods may need to be be re-trained and validated using the training and test sets obtained from the NLST data set. We will improve the nodule segmentation methods so that they can accurately extract the volume for the subsolid, juxta-vascular, and juxta-plural nodules that were found to be difficult in lung nodule segmentation.
2. Design new quantitative radiomic featurs – We have designed several features for differentiating malignant and benign of lung nodules in our previous studies using a limited data set from our own institution. With a large and diverse database like the NLST’s, we will be able to further improve the current features and develop new features to characterize nodule appearances in LDCT images such as shape, volume, textures, locations, and attenuations. We will also design radiomic features to characterize lung parenchyma around lung nodules that may be altered due to a developing malignancy, and to identify and quantify other lung diseases in the lungs. We will design new methods to identify and characterize subsolid nodules, calcified nodules and the distribution pattern of calcification in a nodule. We will develop new features to quantify the attenuation heterogeneity within the ground glass opacity (GGO) nodules identified in the subsolid nodule group. The discriminative power of the radiomic features such as the heterogeneous pattern in differentiating invasive adenocarcinomas from pre-invasive adenocarcinomas , such as adenocarcinoma in situ (AIS) or minimally invasive adenocarcinoma (MIA) will be investigated, using the results of pathological analysis as ground truth.
3. Develop and validate nodule registration methods for temporal change analysis.
4. Develop and validate predictive model using advanced machine learning techniques and statistical methods, with combination of radiomic features and their temporal changes, patient demographic information, and risk factors from medical, occupational or smoking history.
Heang-Ping Chan, PhD, University of Michigan
Lubomir Hadjiiski, PhD,University of Michigan
Jun Wei, PhD,University of Michigan
Ravi Samala, PhD, University of Michigan
Ella Kazerooni, MD, University of Michigan
Aamer Chughtai, MD, University of Michigan
-
Leveraging Serial Low-Dose CT Scans in Radiomics-based Reinforcement Learning to Improve Early Diagnosis of Lung Cancer at Baseline Screening.
Wang Y, Zhou C, Ying L, Lee E, Chan HP, Chughtai A, Hadjiiski LM, Kazerooni EA
Radiol Cardiothorac Imaging. 2024 Jun; Volume 6 (Issue 3): Pages e230196 PUBMED -
Pathologic categorization of lung nodules: Radiomic descriptors of CT attenuation distribution patterns of solid and subsolid nodules in low-dose CT.
Zhou C, Chan HP, Chughtai A, Hadjiiski LM, Kazerooni EA, Wei J
Eur J Radiol. 2020 Aug; Volume 129: Pages 109106 PUBMED