Two-stage lesion detection approach based on dimension-decomposition and 3D context

Lesion detection in Computed Tomography (CT) images is a challenging task the field of computer-aided diagnosis. An important issue to locate area lesion accurately. As branch Convolutional Neural Networks (CNNs), 3D Context-Enhanced (3DCE) frameworks are designed detect lesions on CT scans. The False Positives (FPs) detected 3DCE usually caused by inaccurate region proposals, which slow down inference time. To solve above problems, new method proposed, dimension-decomposition proposal network integrated into framework improve location accuracy detection. Without restriction "anchors" ratios and scales, anchors decomposed independent "anchor strings". Anchor segments dynamically combined accordance with probability, anchor strings different lengths compose bounding boxes. Experiments show that accurate proposals generated our model promote sensitivity FPs spend less time compared current methods.