Nodule Segmentation of Lung CT Images Based on Weakly Supervised Learning

Abstract

Pulmonary nodule segmentation is of great significance in assisting doctors in diagnosis. However, most pulmonary nodules have the characteristics of blurred edges, large differences in individual shapes, and adhesion to the pleural endometrium in CT images. At present, the lung nodule segmentation technology is based on supervised method, and still needs a great deal of manual labeled image data for training. Moreover, the labeling phase is time-consuming, laborious, and subjectively biased, making it difficult to provide large amounts of high-quality image data, which limits the application of deep learning in image segmentation. For the sake of overcome exist problems, this article leads to a segmentation network (WSAK-Net) combining weakly supervised autoencoder with k-means clustering and convolutional network. First, the search starting point is established through the autoencoder, and then the feature search is achieved by k-means clustering. Finally, the semantic lung nodules segmentation is achieved through convolutional networks. Differential analysis of the data by autoencoder and feature explicit processing of weakly labeled data enables the raw data to autonomously generate weakly supervised convergence starting point. Then we use k-means clustering to further extract nodular features, which is based on feature similarity and starts at the convergence starting point, and makes a gradually clustering division of data features. This allows the network to autonomously generate coarse-grained nodule distribution probability labels, and guide the segmentation convolutional network to achieve lung nodule feature extraction and segmentation in weakly supervised mode. We tested our method on our own produced dataset, and the experimental analysis showed that the method can efficiently segment lung nodules in CT images, which is superior to the current state-of-the-art methods.