(C-97) Automated Classification of Lung Cancer Subtypes Using Deep Learning and CT-scan based Radiomic Analysis

Introduction:: Lung cancer is the leading cause of cancer related death in the United States. Computed tomography (CT) imaging remains one of the standards of care diagnostic tools for staging lung cancers. However, conventional interpretation of radiological images can be, to some degree, affected by radiologists’ training and experience and is subjective and mostly qualitative by nature. Finding new solutions for collecting critical microscopic and molecular features with non-invasive and operator-independent approaches remains a high priority in oncology.

Image digitalization coupled with artificial intelligence is emerging as a powerful tool for generating large-scale quantitative data from high-resolution medical images and for identifying patterns that can predict biological processes and pathophysiological changes. Several new tools are now available for image analysis and machine learning processing pipeline allow for automatic segmentation, feature extraction, and model building.

Materials and Methods:: We used previously collected and publicly available CT images retrieved from the Cancer Imaging Archive (TCIA). The TCIA dataset we used (Lung-PET-CT-Dx) contains 251,135 deidentified CT and PET-CT images from lung cancer patients. Radiologist annotations were also provided in the archive along with the images for each CT/PET-CT image.

To perform machine learning-based radiomic analysis, the Computational Environment for Radiologic Research (CERR) software platform was used for deploying the trained incremental Multiple Resolution Residual Network (iMRRN) towards automated segmentation. After direct application of iMMRN segmentation tool, images were visually examined and compared to the manual bounding box annotations.

Using CERR methods, histogram intensity features and tumor morphology features were extracted from the segmented ROIs. These features included seventeen shape features, twenty-two first order features, and eighty texture features from the tumor regions in each CT scan.

To examine the effectiveness of supervised learning methods in classifying lung cancer subtypes using radiomic features extracted from segmented tumor CT data, we trained and tested seven classifiers. Three classes of lung cancer subtypes, namely Adenocarcinoma, Small Cell Carcinoma, and Squamous Cell Carcinoma, were used as response variables for our analysis. The synthetic minority oversampling technique (SMOTE) was used to balance the number of observations in the training dataset. We used chi-square tests to obtain a ranking which describes the degrees of association between each feature and the response variable.


Results, Conclusions, and Discussions:: First, although automated segmentation algorithms such as the iMRRN are designed to operate without prior information about the location of the tumor lesion, we found out that segmentation accuracy is improved when the general location of the tumor is provided. A deep learning (DL) model can then be applied to tumor adjacent slices rather than searching for the tumor in the entire image stack. As this process will only require labeling a single tumor slice, this approach required very limited effort from the radiologist. Thus, it may boost the use of automated DL segmentation and radiomics approached in oncology.
A second important issue that emerged from our analysis is the role of unbalanced dataset as it truly poses challenges in radiomics analysis. When working with retrospective clinical samples, it is common for a dataset to contain unequal number of subjects across groups to be compared. However, it is also known that many machine learning methods are sensitive to unbalanced data as the minority classes would not be sufficiently learned compared to the dominant classes. One should carefully examine the distribution of data before applying machine learning based radiomics analysis because the unsatisfying results could be partially caused by the underrepresentation of some classes. The SVM model had an accuracy of 76.5% and an AUC of 0.86 in the unbalanced comparison versus an accuracy of 92.7% and AUC of 0.97 after the SMOTE function was applied.
This study demonstrated a successful application of the deep learning method iMRRN on segmenting independent lung CT data and identified a procedure to make automatic segmentation more accurate. Direct use of segmentation results using existing deep learning models leads to classification accuracy comparable with published studies. Overall, we demonstrated that classification of subtypes of lung cancer can be semi-automated with minimal radiologist intervention in terms of segmentation and radiomics analysis.


Acknowledgements (Optional): :

References (Optional): :