Fig 1.
Texture features are extracted from the lung parenchyma. Two different MIL classifiers are trained and are tested on previously unseen scans. The results are evaluated against manual annotations performed by two radiologists, a density based analysis, and pulmonary function tests.
Table 1.
Clinical characteristics of subjects belonging to both datasets.
GOLD stratification reflects the classification of the COPD patients according to the GOLD combined risk stratification assessment [3].
Table 2.
miSVM-Q and MILES-Q results using COPD (ClassCOPD) and DLCO (ClassDLCO) labels.
S: separability (×100); AUC: bag AUC (×100).
Fig 2.
2D visualization of patches from COPD and non-COPD subjects using the Gaussian feature representation.
The patches have different distributions, which helps the MIL classifier to classify a subject correctly.
Table 3.
Spearman correlation results with data from pulmonary tests.
ClassCOPD: results from classifier with COPD label; ClassDLCO: results from classifier with DLCO label; Thr LAA: Threshold scan based on low attenuation areas; Agree Exp: area of agreement between the manual annotations of both experts; rho: correlation coefficient.
Fig 3.
Percentage of emphysema (log scale for visibility) per subject annotated by the experts and computed by the classifiers.
Fig 4.
Example of results in randomly selected slices for the density based method, manual annotations from the experts, and classifier results using miSVM-Q and Gaussian features.
From left to right: patients with mild, moderate, severe and very severe COPD.