Papers
Topics
Authors
Recent
Gemini 2.5 Flash
Gemini 2.5 Flash
102 tokens/sec
GPT-4o
59 tokens/sec
Gemini 2.5 Pro Pro
43 tokens/sec
o3 Pro
6 tokens/sec
GPT-4.1 Pro
50 tokens/sec
DeepSeek R1 via Azure Pro
28 tokens/sec
2000 character limit reached

Semi-supervised Medical Image Segmentation through Dual-task Consistency (2009.04448v3)

Published 9 Sep 2020 in cs.CV

Abstract: Deep learning-based semi-supervised learning (SSL) algorithms have led to promising results in medical images segmentation and can alleviate doctors' expensive annotations by leveraging unlabeled data. However, most of the existing SSL algorithms in literature tend to regularize the model training by perturbing networks and/or data. Observing that multi/dual-task learning attends to various levels of information which have inherent prediction perturbation, we ask the question in this work: can we explicitly build task-level regularization rather than implicitly constructing networks- and/or data-level perturbation-and-transformation for SSL? To answer this question, we propose a novel dual-task-consistency semi-supervised framework for the first time. Concretely, we use a dual-task deep network that jointly predicts a pixel-wise segmentation map and a geometry-aware level set representation of the target. The level set representation is converted to an approximated segmentation map through a differentiable task transform layer. Simultaneously, we introduce a dual-task consistency regularization between the level set-derived segmentation maps and directly predicted segmentation maps for both labeled and unlabeled data. Extensive experiments on two public datasets show that our method can largely improve the performance by incorporating the unlabeled data. Meanwhile, our framework outperforms the state-of-the-art semi-supervised medical image segmentation methods. Code is available at: https://github.com/Luoxd1996/DTC

User Edit Pencil Streamline Icon: https://streamlinehq.com
Authors (6)
  1. Xiangde Luo (31 papers)
  2. Jieneng Chen (26 papers)
  3. Tao Song (50 papers)
  4. Yinan Chen (23 papers)
  5. Guotai Wang (67 papers)
  6. Shaoting Zhang (133 papers)
Citations (440)

Summary

Semi-supervised Medical Image Segmentation through Dual-task Consistency

The paper discusses a novel approach to semi-supervised medical image segmentation by leveraging dual-task consistency within a deep learning framework. Medical image segmentation is a crucial task in clinical diagnosis and treatment planning, where the high cost and expertise required for pixel-wise labeling present significant challenges. The proposed framework aims to alleviate these constraints by effectively using unlabeled data alongside a limited amount of labeled data.

Methodology

The authors introduce a dual-task framework consisting of two branches: a pixel-wise segmentation branch and a level set function regression branch. The pixel-wise segmentation branch predicts a pixel-level classification map, while the level set branch captures geometric information by predicting a level set representation of the target. A differentiable task-transform layer is introduced to convert the level set function into a segmentation probability map, enabling consistency regularization between the two tasks. This explicit task-level consistency provides an alternative to the traditional data perturbation strategies found in existing semi-supervised learning (SSL) literature.

The core hypothesis of this research is that enforcing consistency between dual tasks can serve as an effective regularization technique, improving segmentation performance by leveraging unlabeled datasets. The dual-task consistency loss minimizes the difference between the outputs of the pixel-wise classification and the transformed level set function, thus incorporating geometric constraints into the network’s predictions and enabling the utilization of unlabeled data.

Experimental Results

Experiments were conducted on two publicly available 3D medical image datasets: the Left Atrium MRI dataset and the Pancreas CT dataset. The proposed framework significantly outperformed state-of-the-art SSL segmentation methods across multiple evaluation metrics, such as Dice, Jaccard, average surface distance, and 95% Hausdorff Distance. The advantages of the dual-task consistency framework became evident in both fully supervised and semi-supervised settings, where the proposed approach demonstrated robustness and generalization capabilities.

An ablation paper confirmed that the combination of pixel-wise segmentation and level set function regression with task-level consistency significantly enhanced the model's performance. The paper also highlighted the framework’s efficiency in terms of computational cost and training time, making it more practical for real-world applications.

Implications and Future Work

The dual-task consistency framework introduces a promising paradigm for semi-supervised segmentation by highlighting the potential of task-level regularization. This approach opens avenues for further exploration in not only medical imaging but also in other areas of computer vision and image processing, such as multi-task learning, where different but related tasks can enforce consistency constraints to enhance learning from unlabeled data.

In future work, the authors propose exploring additional tasks that can be integrated through differentiable transforms, thus expanding the framework's applicability. Furthermore, they suggest the potential of extending this approach to broader applications, such as video recognition and image reconstruction, to harness the power of unlabeled datasets effectively.

The work presented in this paper contributes significantly to the field of medical image analysis by reducing the dependency on large annotated datasets while maintaining high segmentation accuracy. It challenges conventional SSL methods and paves the way for more sophisticated task-level consistency approaches in various domains.

Github Logo Streamline Icon: https://streamlinehq.com