Efficient Semi-Supervised Gross Target Volume of Nasopharyngeal Carcinoma Segmentation via Uncertainty Rectified Pyramid Consistency (2012.07042v4)

Published 13 Dec 2020 in cs.CV

Abstract: Gross Target Volume (GTV) segmentation plays an irreplaceable role in radiotherapy planning for Nasopharyngeal Carcinoma (NPC). Despite that Convolutional Neural Networks (CNN) have achieved good performance for this task, they rely on a large set of labeled images for training, which is expensive and time-consuming to acquire. In this paper, we propose a novel framework with Uncertainty Rectified Pyramid Consistency (URPC) regularization for semi-supervised NPC GTV segmentation. Concretely, we extend a backbone segmentation network to produce pyramid predictions at different scales. The pyramid predictions network (PPNet) is supervised by the ground truth of labeled images and a multi-scale consistency loss for unlabeled images, motivated by the fact that prediction at different scales for the same input should be similar and consistent. However, due to the different resolution of these predictions, encouraging them to be consistent at each pixel directly has low robustness and may lose some fine details. To address this problem, we further design a novel uncertainty rectifying module to enable the framework to gradually learn from meaningful and reliable consensual regions at different scales. Experimental results on a dataset with 258 NPC MR images showed that with only 10% or 20% images labeled, our method largely improved the segmentation performance by leveraging the unlabeled images, and it also outperformed five state-of-the-art semi-supervised segmentation methods. Moreover, when only 50% images labeled, URPC achieved an average Dice score of 82.74% that was close to fully supervised learning.

View on arXiv

Authors (9)

Xiangde Luo (31 papers)
Wenjun Liao (14 papers)
Jieneng Chen (26 papers)
Tao Song (50 papers)
Yinan Chen (23 papers)
Shichuan Zhang (24 papers)
Nianyong Chen (1 paper)
Guotai Wang (67 papers)
Shaoting Zhang (133 papers)

Citations (184)

View on Semantic Scholar

Summary

Efficient Semi-Supervised Segmentation of Nasopharyngeal Carcinoma Combining Uncertainty Rectification and Pyramid Consistency

The paper by Luo et al., "Efficient Semi-Supervised Gross Target Volume of Nasopharyngeal Carcinoma Segmentation via Uncertainty Rectified Pyramid Consistency," addresses the prevalent challenge of segmenting gross target volume (GTV) in nasopharyngeal carcinoma (NPC) with limited labeled data. The authors propose an innovative framework that combines Uncertainty Rectified Pyramid Consistency (URPC) with a pyramid predictions network (PPNet) to offer a solution that enhances segmentation performance while reducing the requirement for labeled data, leveraging semi-supervised learning (SSL) principles.

In this framework, the PPNet is an extension of a typical CNN, specifically a 3D U-Net, designed to generate multi-scale predictions from the input data. The main novelty lies in the application of URPC for handling unlabeled data. The PPNet generates pyramid predictions, i.e., predictions at multiple scales, and imposes a consistency loss that encourages these predictions to remain consistent, thus learning from unlabeled datasets effectively.

An interesting aspect of the method involves the calculation of uncertainty directly from the variance across pyramid predictions rather than through extensive sampling processes such as Monte Carlo dropout, thus maintaining computational efficiency. This uncertainty estimator is then used to weight the contributions of different predictions in the unsupervised loss function, improving both robustness and reliability of the training process by emphasizing predictions with low uncertainty.

The authors evaluated their framework on a dataset of 258 MRI images of NPC, with only a subset of images labeled. When trained with just 10% and 20% labeled data, the method significantly boosted segmentation performance, achieving dice scores that approached those of fully supervised methods. For instance, the URPC framework achieved an average Dice score of 82.74% with 50% labeled images, closely paralleling the performance under fully supervised conditions.

An important implication of this work is that it provides a viable pathway toward reducing the burden on expert annotators in medical imaging contexts, where obtaining labeled data is often a bottleneck. The URPC framework can generalize to other segmentation tasks beyond NPC, making it a versatile tool in clinical applications.

The research suggests a few avenues for future investigation. For instance, extending the framework to integrate more variational model predictions or optimize the model complexity further can be considered. Additionally, applying the methodology to broader or more diverse datasets can help ascertain its efficiency and effectiveness in varying medical conditions.

In summary, the paper makes a significant contribution to the domain of medical image segmentation by presenting a well-integrated approach combining pyramid predictions, and uncertainty estimation to improve segmentation accuracy in semi-supervised settings. This method stands out by balancing computational demands with improved accuracy, which holds promise for practical applications in medical AI.

PDF Markdown

Related Papers

Find Related Papers