MACCAI LiTS17 Liver Tumor Segmentation Using RetinaNet

Automatic liver tumor detection is an important task and is more challenging in Computed Tomography (CT) due to its size, location, and irregular shape. Detection of small tumors in the liver is problematic because the liver area is mostly covered by the right rib cage. Recent work employs two-stage object detectors that are trained with bounding box annotations. In our work, we apply a simpler and faster one stage detector RetinaNet for localization of liver tumor on LiTS17, and our proposed method precisely detects one or more tumors. This model resolves the class imbalance problem between foreground and background by evaluating the focal loss. Our results on liver tumor CT images from LiTS17 dataset yields strong detection and scored mAP of 96%.