Efficient Semi-supervised Gross Target Volume of Nasopharyngeal Carcinoma Segmentation via Uncertainty Rectified Pyramid Consistency

Xiangde Luo, Wenjun Liao, Jieneng Chen, Tao Song, Yinan Chen, Shichuan Zhang, Nianyong Chen, Guotai Wang, Shaoting Zhang

TL;DR

This paper introduces a novel semi-supervised learning framework called Uncertainty Rectified Pyramid Consistency (URPC) for segmenting Gross Target Volumes (GTV) in Nasopharyngeal Carcinoma (NPC) from MRI scans. URPC efficiently leverages unlabeled data by using multi-scale predictions and an uncertainty-rectifying module, achieving performance close to fully supervised methods with significantly less labeled data.

What Problem Does This Paper Address?

Accurate segmentation of the Gross Target Volume (GTV) in Nasopharyngeal Carcinoma (NPC) is crucial for effective radiotherapy planning. While deep learning models, particularly Convolutional Neural Networks (CNNs), excel at this task, they typically require a large amount of meticulously annotated data. Acquiring these annotations is a time-consuming and expensive process, demanding specialized expertise. This paper addresses the challenge of achieving high-performance GTV segmentation with limited labeled data by proposing a semi-supervised learning approach that effectively utilizes readily available unlabeled medical images.

Methods

The core of the proposed method is the Uncertainty Rectified Pyramid Consistency (URPC) framework. It builds upon a standard 3D UNet backbone, modified to produce Pyramid Predictions (PPNet). This PPNet generates segmentations at multiple scales (resolutions) from a single forward pass.

  1. Pyramid Prediction Network (PPNet): The decoder of the 3D UNet is augmented with prediction layers after each upsampling block. This allows the network to output predictions at different spatial resolutions.
  2. Supervised Loss: For labeled data, a standard supervised loss (combination of Dice and cross-entropy) is applied to these multi-scale predictions.
  3. Pyramid Consistency Loss: For unlabeled data, the key idea is to enforce consistency among the predictions generated at different scales for the same input image. This is achieved by minimizing the discrepancy (variance) between these multi-scale predictions.
  4. Uncertainty Rectifying Module: A crucial innovation is the introduction of an uncertainty estimation mechanism. Instead of relying on computationally expensive methods like Monte Carlo dropout, URPC estimates uncertainty by measuring the KL-divergence between the average prediction across scales and the prediction at each individual scale. This estimation is done efficiently in a single forward pass.
  5. Uncertainty Rectification: The estimated uncertainty is then used to "rectify" the pyramid consistency loss. This means the framework learns to emphasize reliable regions (low uncertainty) and down-weight unreliable or noisy regions (high uncertainty) when enforcing consistency between multi-scale predictions. This is implemented using a voxel- and scale-wise weight derived from the uncertainty, ensuring more stable unsupervised learning.
  6. Overall Loss: The total loss function combines the supervised loss for labeled data and the uncertainty-rectified pyramid consistency loss for unlabeled data, with a time-dependent weighting factor to balance their contributions during training.

Datasets

The study utilizes a dataset of 258 T1-weighted MRI images from patients with Nasopharyngeal Carcinoma (NPC) collected from the West China Hospital. The dataset was split into:

  • Training set: 180 cases
  • Validation set: 20 cases
  • Testing set: 58 cases

For the semi-supervised experiments, a small fraction of the training data was used as labeled:

  • 10% labeled data: 18 images (out of 180) were labeled, with the remaining 162 used as unlabeled.
  • 20% labeled data: Used for further comparison in later experiments.

The paper does not introduce a new benchmark dataset but rather uses a clinical dataset for evaluation.

Key Insights and Ideas

  • Multi-scale consistency is powerful: The core insight is that encouraging predictions from the same network at different resolutions to agree can act as a strong regularizer for semi-supervised learning, even without explicit task-specific knowledge.
  • Uncertainty estimation from prediction discrepancy is efficient: Estimating model uncertainty by looking at how much predictions vary across different scales (within a single forward pass) is a computationally efficient alternative to methods requiring multiple passes.
  • Uncertainty guides learning: Using uncertainty to selectively focus on reliable regions during consistency training is key to preventing noise from unlabeled data from degrading performance. This "rectification" makes the learning process more robust.
  • The sweet spot for scales: The study found that while more scales generally improve performance, there's a point of diminishing returns (S=4 was optimal in their experiments), likely because very low-resolution scales might lose too much detail.

Key Results

  • Significant improvement with limited labels: With only 10% labeled data, URPC significantly improved segmentation performance compared to a baseline supervised model trained on the same limited data.
  • Outperforms state-of-the-art: URPC outperformed five other state-of-the-art semi-supervised segmentation methods when using 10% labeled data.
  • Near fully supervised performance: When using 50% of the training images as labeled data, URPC achieved a mean Dice score of 82.74%, which is very close to the performance of a fully supervised model (83.51%).
  • Efficiency: The method requires only a single forward pass per image for uncertainty estimation, making it more efficient than methods requiring multiple forward passes.
  • Qualitative results: Visualizations show that URPC produces segmentations with higher overlap and fewer false negatives compared to baselines, with uncertainty maps highlighting boundaries as regions of higher uncertainty.

Limitations and Future Work

  • Limitations:
    • The paper notes that while S=4 was optimal, S=5 showed slightly lower performance, suggesting that very low-resolution scales might not always be beneficial.
  • The method's efficiency is highlighted, but the exact computational cost and GPU memory requirements compared to other methods are discussed in terms of "more efficient" rather than precise figures for all scenarios.
  • Future Work:
    • The authors plan to evaluate the URPC framework on other segmentation tasks beyond NPC GTV segmentation, indicating its potential for broader applicability.
  • They suggest that the framework's ability to leverage multi-scale outputs and uncertainty estimation could be applied to other areas of medical image analysis.

Summary updated

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% labeled images, URPC achieved an average Dice score of 82.74% that was close to fully supervised learning. Code is available at: https://github.com/HiLab-git/SSL4MIS .
BibTeX
@inproceedings{luo2021efficient,
  author = {Xiangde Luo and Wenjun Liao and Jieneng Chen and Tao Song and Yinan Chen and Shichuan Zhang and Nianyong Chen and Guotai Wang and Shaoting Zhang},
  title = {{Efficient Semi-supervised Gross Target Volume of Nasopharyngeal Carcinoma Segmentation via Uncertainty Rectified Pyramid Consistency}},
  booktitle = {International Conference on Medical Image Computing and Computer-Assisted Intervention},
  year = {2021},
  url = {https://openreview.net/forum?id=b1PhH7BEQpp},
}
Figure