An adaptive multi-scale network with nonorthogonal multi-union input for reducing false positive of lymph nodes

Abstract

Lymph node (LN) detection and analysis are of great significance in terms of cancer staging and measuring the effectiveness of treatment. But, it is still a challenging and laborious task due to the lack of adequately-labeled data and the similar pathological features with surrounding structures in computed tomography (CT). In this paper, we propose a new representation for lymph node detection after augmentation, which can effectively decompose candidate cubic CT images by generating nonorthogonal multi-union 2D slices. These new views with coupling relationship will be used as a novel input to train the convolutional neural networks (CNNs) to achieve the purpose of reducing false positives (FP). In order to further adapt to the mutative radiological characteristics of lymph nodes, we designed an adaptive multi-scale network. This model adaptively learns the features of different scales images and redistributes weights of convolution kernel to optimize the classification result. We validate the approach on two datasets 90 CT volumes with 388 mediastinal LNs and 86 patients with 595 abdominal LNs. Our proposed methods perform better in both two cases, sensitivities reached 78%, 86% at 3 false-positives per patient volume (FP/vol.), and 94%, 96% at 6FP/vol. in mediastinum and abdomen respectively.