Automatic segmentation of organs at risk and tumors in CT images of lung cancer from partially labelled datasets with a semi-supervised conditional nnU-Net. (November 2021)

• Record Type: Journal Article
• Authors: Zhang, Guobin; Yang, Zhiyong; Huo, Bin; Chai, Shude; Jiang, Shan
• Is Part Of: Computer methods and programs in biomedicine. Volume 211(2021)
• Publication Date: 2021-11
• Subjects: Automatic segmentation -- Partially labelled dataset -- Conditioning strategy -- Semi-supervised learning -- Deep learning; Medicine -- Computer programs -- Periodicals; Biology -- Computer programs -- Periodicals; Computers -- Periodicals; Medicine -- Periodicals; Médecine -- Logiciels -- Périodiques; Biologie -- Logiciels -- Périodiques; Biology -- Computer programs; Medicine -- Computer programs; Periodicals; Electronic journals; 610.28
• DOI: 10.1016/j.cmpb.2021.106419
• Languages: English
• ISSNs: 0169-2607

Abstract

Highlights: A semi-supervised conditional nnU-Net was proposed to segment lung tumors and OARs. The model was trained on partially labelled datasets using a conditioning strategy. Semi-supervised learning for multiclass segmentation is applied. Two combined loss functions are individually designed for OARs and tumor segmentation. This method has the potential to help radiologists with RT treatment planning. Abstract: Background and Objective: Accurately and reliably defining organs at risk (OARs) and tumors are the cornerstone of radiation therapy (RT) treatment planning for lung cancer. Almost all segmentation networks based on deep learning techniques rely on fully annotated data with strong supervision. However, existing public imaging datasets encountered in the RT domain frequently include singly labelled tumors or partially labelled organs because annotating full OARs and tumors in CT images is both rigorous and tedious. To utilize labelled data from different sources, we proposed a dual-path semi-supervised conditional nnU-Net for OARs and tumor segmentation that is trained on a union of partially labelled datasets. Methods: The framework employs the nnU-Net as the base model and introduces a conditioning strategy by incorporating auxiliary information as an additional input layer into the decoder. The conditional nnU-Net efficiently leverages prior conditional information to classify the target class at the pixelwise level. Specifically, we employ the uncertainty-aware mean teacher (UA-MT) framework to assist in OARs segmentation, which can effectively leverage unlabelled data (images from a tumor labelled dataset) by encouraging consistent predictions of the same input under different perturbations. Furthermore, we individually design different combinations of loss functions to optimize the segmentation of OARs (Dice loss and cross-entropy loss) and tumors (Dice loss and focal loss) in a dual path. Results: The proposed method is evaluated on two publicly available datasets of the spinal cord, left and right lung, heart, esophagus, and lung tumor, in which satisfactory segmentation performance has been achieved in term of both the region-based Dice similarity coefficient (DSC) and the boundary-based Hausdorff distance (HD). Conclusions: The proposed semi-supervised conditional nnU-Net breaks down the barriers between nonoverlapping labelled datasets and further alleviates the problem of "data hunger" and "data waste" in multi-class segmentation. The method has the potential to help radiologists with RT treatment planning in clinical practice.