Local Characteristic Features for Computer Aided Detection of Pulmonary Embolism in CT Angiography

An automated detection system is constructed for detecting pulmonary embolism from computed tomographic pulmonary angiographic images. Our previous work has presented novel effective algorithms to identify suspicious PE regions from images and reduce false detections by designing powerful classifiers. However, these techniques have to take effects in conjunction with discriminative features used to characterize each identified PE candidate. This paper investigates three sets of novel features: 1. features based on local candidate co-occurrence matrices to remove false detections due to noise and poorly mixed contrast; 2. features characterizing vessel properties to eliminate candidates outside of vessel; 3. features discriminating between arteries and veins to remove candidates from veins. We tested these features in our multiple instance learning classification setting, and they constantly demonstrated performance improvement when the 3 sets of features are included sequentially. The resulted PE CAD system has capabilities of incrementally reporting any detection immediately once becoming evident during searching, offering real-time support and achieving 85% sensitivity at 5 false positives.