Multi-task weakly-supervised learning model for pulmonary nodules segmentation and detection

Abstract

For two-dimensional (2D) continuity characteristics of pulmonary nodules CT images, a sequence segmentation model based on U-shaped structure network and Convolutional Long Short-Term Memory (ConvLSTM) network is proposed to fully obtain the context space characteristics of image slices. In order to solve the problem of limited number of annotated samples in pulmonary nodules segmentation task, a segmentation method based on multi-task learning framework is proposed, which uses the annotated data of different types of tasks to mine the potential common characteristics among tasks; aiming at the problem of unbalanced category distribution in pulmonary nodules segmentation task, the design method of unified loss function under the multi-task learning framework is studied, and an optimization strategy integrating image prior knowledge and dynamic adjustment of multi-task weight is proposed to ensure that each task can complete training and learning efficiently. The experiments based on the LIDC-IDRI dataset demonstrate that the multi-task learning method proposed in this paper for the segmentation of pulmonary nodules under weak supervision is optimized from the three aspects of model design, network structure and constraints, and the MIoU and DSC are improved to 79.23% and 82.26% respectively.