A method for automatically quantifying emphysema regions usingHigh-Resolution Computed Tomography (HRCT) scans of patients with chronicobstructive pulmonary disease (COPD) that does not require manually annotatedscans for training is presented. HRCT scans of controls and of COPD patientswith diverse disease severity are acquired at two different centers. Texturalfeatures from co-occurrence matrices and Gaussian filter banks are used tocharacterize the lung parenchyma in the scans. Two robust versions of multipleinstance learning (MIL) classifiers, miSVM and MILES, are investigated. Theclassifiers are trained with the weak labels extracted from the forcedexpiratory volume in one minute (FEV$_1$) and diffusing capacity of the lungsfor carbon monoxide (DLCO). At test time, the classifiers output a patientlabel indicating overall COPD diagnosis and local labels indicating thepresence of emphysema. The classifier performance is compared with manualannotations by two radiologists, a classical density based method, andpulmonary function tests (PFTs). The miSVM classifier performed better thanMILES on both patient and emphysema classification. The classifier has astronger correlation with PFT than the density based method, the percentage ofemphysema in the intersection of annotations from both radiologists, and thepercentage of emphysema annotated by one of the radiologists. The correlationbetween the classifier and the PFT is only outperformed by the secondradiologist. The method is therefore promising for facilitating assessment ofemphysema and reducing inter-observer variability.
展开▼
