Annotated Lung CT Image Database

Computed tomography (CT) of lungs provides a diagnostic tool for identifying a range of lesions and diseases visible in the obtained scans. For helping radiologists in a timely and efficient assessment of a large number of scans different machine learning methods have been applied for the detection and classification of abnormalities. However, before clinical usage of such algorithms it is necessary to achieve high accuracy of the algorithm. This is achieved through training and testing phases for which it is essential to have a database that would include a range of abnormalities in the lungs. Here we propose a novel database ALCTID (Annotated Lung CT Image Database) with regions of interest (ROI) annotated by an experienced thoracic radiologist. Database includes 170 lung CT images with a total of 307 annotated ROIs comprising a range of abnormalities, from cancerous lesions, enlarged lymph nodes to enlarged heart and edema. To demonstrate the applicability of the novel database we trained and tested convolution neural network based on the YOLO (You Only Look Once) algorithm on 170 images with annotated ROIs and 170 images of healthy lungs. Out of 80 annotated ROIs in the test set, the network correctly detected 56 ROIs, with 12 false positives and 25 false negatives. The new database ALCTID is publicly available at http://www.vcl.fer.hr/alctid.