Evaluating Deep Learning Algorithms in Pulmonary Nodule Detection

Lung cancer is considered the deadliest cancer worldwide. In order to detect it, radiologists need to inspect multiple Computed Tomography (CT) scans. This task is tedious and time consuming. In recent years, promising methods based on deep learning object detection algorithms were proposed for the automatic nodule detection and classification. With those techniques, Computed Aided Detection (CAD) software can be developed to alleviate radiologist’s burden and help speed-up the screening process. However, among available object detection frameworks, there are just a limited number that have been used for this purpose. Moreover, it can be challenging to know which one to choose as a baseline for the development of a new application for this task. Hence, in this work we propose a benchmark of recent state-of-the-art deep learning detectors such as Faster-RCNN, YOLO, SSD, RetinaNet and EfficientDet in the challenging task of pulmonary nodule detection. Evaluation is done using automatically segmented 2D images extracted from volumetric chest CT scans.