PaperlayerView exact revision on arXiv

Source-linked AI summary

Double-Uncertainty Weighted Method for Semi-supervised Learning

Yixin Wang, Yao Zhang, Jiang Tian, Cheng Zhong, Zhongchao Shi, Yang Zhang, Zhiqiang He

arXiv:2010.09298v1cs.CV

TL;DR

Reliable labels for medical image segmentation are costly, and unlabeled teacher predictions lack ground-truth verification. This paper uses segmentation and feature uncertainty in a teacher-student method to guide consistency training, and reports better performance than uncertainty-based semi-supervised methods on two medical datasets.

  • Problem

    Expert-labeled medical images are expensive, while unlabeled data provide no ground truth for judging whether teacher predictions are accurate.

    Source: §1 · p. 2

  • Method

    A teacher-student segmentation method uses Monte-Carlo dropout to estimate segmentation and feature uncertainty, adjusts teacher predictions through a learnable consistency loss, and weights consistency penalties by uncertainty.

    Source: §1 · p. 2, §2.2 · p. 4, Learnable uncertainty consistency loss · p. 5

  • Results

    The method outperforms state-of-the-art uncertainty-based semi-supervised segmentation approaches on two medical datasets.

    Source: §1 · p. 3, §4 · p. 9

  • Takeaways & Limitations

    The paper concludes that double uncertainty supports semi-supervised segmentation with limited labeled images across the two evaluated medical datasets.

    Source: §4 · p. 9

Abstract

from arXiv · show

Though deep learning has achieved advanced performance recently, it remains a challenging task in the field of medical imaging, as obtaining reliable labeled training data is time-consuming and expensive. In this paper, we propose a double-uncertainty weighted method for semi-supervised segmentation based on the teacher-student model. The teacher model provides guidance for the student model by penalizing their inconsistent prediction on both labeled and unlabeled data. We train the teacher model using Bayesian deep learning to obtain double-uncertainty, i.e. segmentation uncertainty and feature uncertainty. It is the first to extend segmentation uncertainty estimation to feature uncertainty, which reveals the capability to capture information among channels. A learnable uncertainty consistency loss is designed for the unsupervised learning process in an interactive manner between prediction and uncertainty. With no ground-truth for supervision, it can still incentivize more accurate teacher's predictions and facilitate the model to reduce uncertain estimations. Furthermore, our proposed double-uncertainty serves as a weight on each inconsistency penalty to balance and harmonize supervised and unsupervised training processes. We validate the proposed feature uncertainty and loss function through qualitative and quantitative analyses. Experimental results show that our method outperforms the state-of-the-art uncertainty-based semi-supervised methods on two public medical datasets.

1 Introduction

Medical segmentation depends on costly expert-labeled data, and unlabeled teacher predictions are difficult to verify. The paper proposes a teacher-student method using segmentation and feature uncertainty to guide training, and reports outperforming uncertainty-based semi-supervised methods on two medical datasets.

1 Introduction

  • High-quality medical annotations are costly because they require domain expertise, motivating semi-supervised learning with unlabeled images.Teacher-student methods penalize inconsistent teacher and student predictions, but unlabeled data lack ground truth for checking teacher accuracy.Source: §1 · p. 2
  • Earlier uncertainty-guided methods use manually set thresholds, which can miss incorrect low-uncertainty predictions or filter out useful teacher guidance.The paper motivates uncertainty-weighted training as an alternative to threshold-based information flow.Source: §1 · p. 2
  • The teacher uses Monte-Carlo dropout to estimate segmentation and feature uncertainty, which modifies its predictions and weights student consistency training.The student combines supervised segmentation loss with consistency loss on labeled and unlabeled data.Source: §1 · p. 2, Figure 1 · p. 3
  • The learnable uncertainty consistency loss is designed to improve teacher predictions and reduce uncertainty without ground-truth supervision.Double uncertainty also assigns smaller loss weights to less trustworthy predictions.Source: §1 · p. 2
  • The method outperforms state-of-the-art uncertainty-based semi-supervised segmentation approaches on the left-atrium and kidney datasets.The introduction identifies the 2018 Atrial Segmentation Challenge and MICCAI 2019 KiTS Challenge.Source: §1 · p. 3

2 Method

The method extends teacher–student semi-supervised segmentation with feature and segmentation uncertainty. It uses uncertainty to adjust teacher guidance and weight consistency training alongside supervised segmentation loss.

2.1 Feature uncertainty

  • Feature uncertainty is estimated from channel feature maps across stochastic inference passes, with relative channel uncertainty indicating information-capture ability.The paper treats each channel as a feature detector and the overall estimate as an indication of model feature-extraction ability.Source: §2.1 · pp. 3–4

2.2 Double-uncertainty weighted model

  • The student minimizes supervised segmentation loss on labeled data and consistency loss on both labeled and unlabeled data, with the teacher providing guidance.The teacher uses exponential moving average weights of the student; dropout and random input noise regularize teacher training.Source: §2.2 · p. 4
  • Segmentation uncertainty is estimated from stochastic teacher predictions, and voxel-wise uncertainty adjusts teacher probabilities by interpolating them with student predictions.When teacher uncertainty is high, the adjusted prediction moves toward the student prediction; with low uncertainty, teacher guidance remains unchanged.Source: Learnable uncertainty consistency loss · p. 5
  • The total objective combines supervised segmentation loss with the proposed consistency loss, while double uncertainty scales consistency penalties for individual predictions.Feature and segmentation uncertainty lower the consistency weight when teacher predictions are highly uncertain; a log function restricts the effect of extremely small segmentation uncertainty.Source: Uncertainty weight · pp. 5–6

3 Experiment

The experiments compare the approach on left atrium and kidney segmentation under limited labels, and examine feature uncertainty and the proposed loss and uncertainty weight. The reported comparisons favor the method on both tasks, while validation finds feature uncertainty decreases during training.

3 Experiment

  • Evaluation covers 100 LA scans and 210 kidney scans, with 80/20 and 160/50 train/test splits, respectively, using Dice, Jaccard, 95HD, and ASD.Training uses dropout for 16 uncertainty estimates and estimates feature uncertainty at the fourth upsampling layer.Source: §3 · p. 6, LA segmentation · p. 6, Kidney segmentation · pp. 6–7

3.1 Implementation details

  • With 16 or 8 labeled LA scans, the method outperforms UA-MT on all four indices; with 16, it approaches V-Net trained on all 80.Table 1 compares supervised and semi-supervised methods.Source: LA segmentation · p. 6, Table 1 · p. 7
  • Using 4 labeled kidney scans, the method ranks top among compared methods and improves Dice most when labeled data are scarce.Figure 2(a) compares Dice results across labeled and unlabeled data amounts.Source: Figure 2 · p. 7, Kidney segmentation · p. 7

3.2 Uncertainty validation

  • Feature uncertainty stays relatively ordered across channels while declining over training; maps concentrate uncertainty near targets, especially boundaries and tissue joints.The figure’s brighter regions represent greater uncertainty.Source: §3.2 · p. 8, Figure 3 · p. 8
  • The learnable consistency loss improves teacher predictions and reduces teacher uncertainty; uncertainty weights also improve UA-MT.Table 2 covers the contributions of the consistency loss and uncertainty weight.Source: §3.2 · p. 8, Table 2 · p. 9

4 Conclusion

The conclusion presents uncertainty as both a teacher-training signal and a consistency weight in semi-supervised learning. Across two medical datasets, the authors report improvements using limited labeled images.

  • The paper introduces feature uncertainty and evaluates it as part of using uncertainty estimates in a teacher-student model for semi-supervised learning.Source: §4 · p. 9
  • The learnable uncertainty consistency loss facilitates a more accurate teacher model with less uncertainty.Source: §4 · p. 9
  • Uncertainty serves as a consistency weight to balance unsupervised training, and experiments on two medical datasets report improvements with limited labeled images.Source: §4 · p. 9
Loading 2010.09298v1…