Keywords:
Cancer, Staging, CT, Thorax, Mediastinum
Authors:
C. Blüthgen1, M. Patella2, I. Schmitt-Opitz1, W. Weder2, T. Frauenfelder2; 1Zürich/CH, 2Zurich/CH
DOI:
10.26044/esti2019/P-0110
Methods & Materials
In this retrospective study,
CT scans of patients with histologically confirmed and categorized thymic neoplasms were extracted from the PACS and reformatted to a fixed slice thickness of 2 mm and an in-plane resolution of 1 x 1 mm².
Tumors were manually segmented using the open source software 3D Slicer[2].
Either using the original image or applying filters (Laplace of Gaussian filter,
Wavelet filter),
pre-defined,
3D-based texture features of 7 major categories were calculated (First Order Statistics,
Shape-based Features,
Gray Level Cooccurence Matrix,
Gray Level Run Length Matrix,
Gray Level Size Zone Matrix,
Neigbouring Gray Tone Difference Matrix and Gray Level Dependence Matrix). Texture features were calculated using the python library pyradiomics[3].
As small differences in segmentation may influence calculated features,
volume of interest (VOI) placements were repeated by the original reader after 2 weeks,
as well as by a second radiologist.
Intra- and interreader agreement was assessed using the intraclass correlation coefficient (ICC(3,1)[4]).
Features with an ICC lower than 0.61 (indicating poor to moderate agreement) were excluded.
Correlation-based feature subset selection was performed using an open-source data mining software (WEKA[5]).
This method evaluates the worth of a subset of features by considering the individual predictive ability of each feature along with the degree of redundancy between the set,
thereby favoring features that are highly correlated with the class while having low inter-correlation.
The selected features for each staging system were used to train a machine learning classifier.
The meta-algorithm AdaBoost[6] iteratively tweaks a combination of several base classifiers to improve on previously misclassified training instances.
The final classification results from a weighted vote of the underlying classifiers.
Random Forest classifiers[7] were used as base learners.
Classification performance was estimated using 10-fold cross-validation and expressed as accuracy as well as area under the receiver operator characteristic curve (ROC AUC).