Improved Focus on Hard Samples for Lung Nodule Detection

TL;DR

This work presents an improved detection network that pays more attention to hard samples and datasets to deal with lung nodules by introducing deformable convolution and self-paced learning.

Abstract

Recently, lung nodule detection methods based on deep learning have shown excellent performance in the medical image processing field. Considering that only a few public lung datasets are available and lung nodules are more difficult to detect in CT images than in natural images, the existing methods face many bottlenecks when detecting lung nodules, especially hard ones in CT images. In order to solve these problems, we plan to enhance the focus of our network. In this work, we present an improved detection network that pays more attention to hard samples and datasets to deal with lung nodules by introducing deformable convolution and self-paced learning. Experiments on the LUNA16 dataset demonstrate the effectiveness of our proposed components and show that our method has reached competitive performance.

Figures (4)

• Figure 1: (a) shows an indication of fixed receptive fields in standard convolution (up) and adaptive receptive fields in deformable convolution (down). In (b), the left of each image triplet shows the sampling location of a $3\times 3$ filter on the preceding feature map of lung nodules. Curriculum learning strategy simulates the learning process of human beings for new knowledge, dividing the training process into several stages from the simple to the difficult, and eventually mastering it.
• Figure 2: The deformable convolution operation.
• Figure 3: The pipeline of our method. (a) shows the architecture of the improved U-Net, and (b) shows the multi-branch ensemble learning architecture based on three kinds of 3D CNN models that use the dual pooling approach. The CT images are processed by the improved U-Net that can detect all lung nodule candidates, and then the multi-branch ensemble learning 3D CNN utilizes the Self-Paced Learning strategy to reduce the false positive lung nodules. In the figure, "Max Pool" means an adaptive max pooling layer with the output size, "Up-Conv" means a standard convolutional layer with the kernel size, "DConv" means a deformable convolutional layer with the kernel size and output channel number, "BatchNorm" means batch normalization operation, "DB" defines a dense block, "DDB" defines a deformable dense block, "Deformable RDB" defines a deformable residual dense block, "3DDP-DenseNet" defines a 3D-DenseNet based on dual pooling (DP) and the other two are similarly defined.
• Figure 4: The detection results of our method.