Automatic Detection of Focal Liver Lesions in Multi-phase CT Images Using One-Stage Object Detectors

Multi-phase CT images are widely used for diagnosis of focal liver lesions. Automatic detection of focal liver lesions in multi-phase CT images is an essential step in computer-aided diagnosis. In recent years, many deep learning methods have been proposed for object detection tasks and have achieved the state-of-the-art performance. In this paper, we study the detection performance of using one-stage object detection framework (i.e., YOLOv3, YOLOv5 and YOLOv7) to detect focal liver lesions in CT images. We also compare the performance with two-stage detection frameworks. We analyze detection accuracy of focal liver lesions from different phases of CT scan (i.e., non-contract enhanced (NC) phase, arterial (ART) phase, and portal venous (PV) phase). According to the analysis we can determine the CT phase which is best suited to locate the lesion. Experimental results show that the one-stage object detectors achieve state-of-the-art performance while the PV phase is most appropriate for lesion detection.