
Contact the VUNO team anytime.
RSNA 2018
Purpose To ensure generalization performance in various hospitals, we developed a deep learning based automatic Chest PA screening System which can detect 5 class findings and performs well on various devices. Its performance was evaluated by using FROC and FOM in various devices.
Method and Materials Our system can detect 5 class findings which are nodule, consolidation, pleural effusion, interstitial opacity, pneumothorax. We used chest radiographs(PA view) collected from 2013 to 2015 at two hospitals. There were 18,869 labeled CRs comprised of 11,181 normal and 1,943 nodule, 1,293 consolidation, 1,867 pleural effusion, 1,406 interstitial opacity, 1,179 pneumothorax CRs. All abnormal CRs were clinically confirmed by CT scans. Thus, we could collected elaborate segmentation label data. Furthermore, our data were acquired in various devices, the first hospital has x-ray detectors from 3 manufacturers( GE, FUJI, Canon ), and the other has devices from 2 manufacturers( Philips, LISTEM ). We used only data from the first hospital for training and then used all data of other hospital as test data.
Results In the test dataset, our screening performance showed AUC of 0.99, with an sensitivity, specificity of 97.6%, 97.9%, respectively. For each class findings, Our system achieved FROC and FP per scan of 81.2% / 0.72, 84.4% / 1.41, 83.1% / 0.35, 85.7% / 1.45, 91.6% / 0.78 for nodule, consolidation, pleural effusion, interstitial opacity, pneumothorax, respectively. In FOM performance, sensitivity is 84.7%, 91.5%, 93.7%, 97.9%, 97.2%.
Conclusion Our screening system demonstrated reliable performance in various devices and multi hospitals. It also showed competitive results in detecting location and classification of 5 class findings.
Clinical Relevance/Application Our system demonstrates reliable performance on various devices on multiple findings using CAD.