[Paper Review] Semi-supervised Anatomical Landmark Detection via Shape-regulated Self-training

Runnan Chen, Yuexin Ma+5 more|arXiv (Cornell University)|May 28, 2021

Medical Imaging and Analysis35 references4 citations

TL;DR

This paper proposes a shape-regulated self-training framework for semi-supervised anatomical landmark detection that leverages a PCA-based shape prior to refine inconsistent pseudo-labels and a Region Attention loss to guide the network toward structurally consistent regions. The method achieves state-of-the-art performance, improving mean registration error by 6.3% on benchmark datasets and showing strong generalization with minimal labeled data.

ABSTRACT

Well-annotated medical images are costly and sometimes even impossible to acquire, hindering landmark detection accuracy to some extent. Semi-supervised learning alleviates the reliance on large-scale annotated data by exploiting the unlabeled data to understand the population structure of anatomical landmarks. The global shape constraint is the inherent property of anatomical landmarks that provides valuable guidance for more consistent pseudo labelling of the unlabeled data, which is ignored in the previously semi-supervised methods. In this paper, we propose a model-agnostic shape-regulated self-training framework for semi-supervised landmark detection by fully considering the global shape constraint. Specifically, to ensure pseudo labels are reliable and consistent, a PCA-based shape model adjusts pseudo labels and eliminate abnormal ones. A novel Region Attention loss to make the network automatically focus on the structure consistent regions around pseudo labels. Extensive experiments show that our approach outperforms other semi-supervised methods and achieves notable improvement on three medical image datasets. Moreover, our framework is flexible and can be used as a plug-and-play module integrated into most supervised methods to improve performance further.

Motivation & Objective

• Address the high cost and expertise required for manual annotation of anatomical landmarks in medical imaging.
• Reduce reliance on large-scale fully annotated datasets by leveraging unlabeled data in semi-supervised learning.
• Improve pseudo-label consistency in self-training by incorporating global anatomical shape constraints often ignored in prior methods.
• Develop a flexible, plug-and-play module compatible with existing heatmap-based supervised methods to boost performance.

Proposed method

• Use a pre-trained backbone network to generate initial pseudo-labels for unlabeled data.
• Apply a PCA-based shape model to regularize pseudo-labels by projecting them onto a learned shape space and filtering out abnormal or inconsistent predictions.
• Introduce a Region Attention loss that encourages the network to focus on spatially consistent structural regions around pseudo-labeled landmarks, rather than forcing fixation on exact points.
• Perform iterative self-training: refine pseudo-labels via shape regulation, optimize with Region Attention loss, and update the network to better predict ground truth.
• Ensure robustness across different imaging devices by training on unlabeled data from diverse sources.
• Integrate the framework as a plug-and-play module into existing supervised landmark detection models.

Experimental results

Research questions

• RQ1 Can global anatomical shape constraints be effectively leveraged to improve pseudo-label quality in semi-supervised landmark detection?
• RQ2 How does shape regulation via PCA-based prior reduce inconsistency in pseudo-labels during self-training?
• RQ3 To what extent does the Region Attention loss enhance feature learning by focusing on structurally coherent regions around landmarks?
• RQ4 How does the method generalize when unlabeled data comes from different imaging devices than the labeled data?
• RQ5 What performance gain can be achieved with minimal labeled data using this shape-regulated self-training approach?

Key findings

• The proposed method achieves a relative improvement of 6.3% in mean registration error on the 3D skull dataset compared to existing semi-supervised methods.
• On the cephalogram dataset, the method achieves a 16.7% improvement with only 20 labeled samples, demonstrating strong data efficiency.
• Abllation studies show that removing the Region Attention loss increases error from 0.8925mm to 0.9293mm, confirming its critical role in performance.
• Without shape regulation, the error increases from 0.8925mm to 0.9117mm, indicating that shape prior stabilizes self-training and improves accuracy.
• Performance improves with more unlabeled data: gains of 4.4%, 5.8%, and 6.7% were observed with 150, 300, and 600 unlabeled samples, respectively.
• The method generalizes well across different imaging devices, maintaining high performance even when unlabeled data is from a different scanner than the labeled data.