Multi-context 3D Resnet for Small-size False Positive Reduction in Pelvic Lymph Node Detection

Zhen Pan, Shuo Huang, Han Wang, Mingtian Wei, Junjie Cui, Ziqiang Wang, Haixian Zhang

Abstract: False positive reduction(FPR) plays a crucial role in abdominal lymph node detection system, which is of great significance in colorectal cancer diagnosis and early treatment. However, this remains a challenge owing to the complexity of the abdominal tissue. In this study, a simple yet effective method for FPR in the small-size abdominal lymph nodes detection is proposed. For small-size lymph nodes, we design a 3D residual network to adapt to corresponding input, and conveniently adjust the network structure and parameters according to the amount of data. Moreover, we use a multi-context fusion method to integrate the results of multiple models to meet the challenge of vary in lymph node volume. Due to the open-source CTLNDataset with only large lymph node, we utilize a new dataset named PLN-Dataset, which contains a large number of small-size pelvic lymph nodes. The proposed method gets an area under curve(AUC) value of 0.991 in PLNDataset, a good performance metric on the competition performance metric(CPM) with a score of 0.837, and also achieve competitive results in CTLNDataset. The result shows that the proposed method is effective and robust for small-size abdominal lymph nodes.