Abstract
Few-shot medical landmark detection relies heavily on the quality of selected template images. However, the impact of template selection is often overlooked, leading to significant performance variance. This paper introduces the Sample Choosing Policy (SCP), a self-supervised framework designed to identify the most representative images in an unlabeled dataset to serve as templates. By leveraging pixel-wise contrastive learning and SIFT-based keypoint proxies, SCP identifies the "center" of the data distribution. Experimental results on cephalometric and hand X-ray datasets demonstrate that SCP significantly reduces Mean Radial Error (MRE) compared to random selection and traditional active learning, effectively bridging the gap between few-shot and fully supervised models.
Introduction
Deep learning in medical imaging typically demands massive, expert-annotated datasets. Few-shot learning offers a solution by enabling models to detect landmarks—such as jaw corners or bone joints—using only a handful of labeled "template" images. However, a critical observation drives this research: not all templates are equally effective. Experiments show that selecting the "best" image as a template can yield a Mean Radial Error (MRE) of 2.9mm, while the "worst" results in 4.5mm. This performance gap, dictated by the luck of initial selection, presents a major hurdle for reliable clinical deployment. The challenge lies in identifying "worthy" images for annotation before any ground-truth labels exist.
Methods: The Sample Choosing Policy (SCP)
The authors propose the Sample Choosing Policy (SCP), a three-stage self-supervised framework to identify the most representative images in an unlabeled pool.
- Self-Supervised Feature Extraction: The model is trained on all unlabeled images using a pixel-wise contrastive learning task (InfoNCE loss). This allows the model to learn local patterns and textures without explicit labels.
- Key Point Proposal: To proxy anatomical landmarks, the system uses the Scale-Invariant Feature Transform (SIFT) algorithm to identify "interesting" key points. These serve as structural anchors for the subsequent evaluation.
- Representative Score Estimation: The model calculates similarity scores across the dataset, seeking the image whose keypoint features are most central to the overall data distribution in the latent space.
Key Insights: Representativeness vs. Uncertainty
A core contribution of this work is the shift in selection philosophy. Traditional Active Learning (AL) strategies typically prioritize "hard" or "unfamiliar" examples to expand the model's decision boundary. In contrast, for few-shot templates, the authors argue that "representativeness" is paramount.
By selecting the most "average" or central image, the model can more easily map features to the widest variety of other cases. Furthermore, the study found that SIFT key points, while not identical to anatomical landmarks, are sufficiently co-located to serve as effective proxies for identifying representative anatomical features.
Results and Performance
The SCP was evaluated on Cephalometric X-ray, Hand X-ray, and WFLW (facial) datasets. Key findings include:
- Error Reduction: On the Hand X-ray dataset, SCP reduced MRE from 4.114mm (random) to 2.635mm—a 35.5% improvement.
- State-of-the-Art Performance: A 5-shot model using SCP-selected templates achieved an MRE of 2.33mm, approaching the performance of fully supervised models.
- Reliability: SCP effectively "raises the floor" of performance; even its worst-case selections outperformed the average case of random selection, significantly reducing the risk of poor model initialization.
Conclusion
The Sample Choosing Policy demonstrates that the "which to label" question is as important as "how to train" in few-shot medical imaging. By prioritizing representativeness through self-supervision, SCP provides a robust, label-free method for optimizing landmark detection. Future work will focus on improving the efficiency of multi-template selection and extending the policy to other medical imaging tasks like organ segmentation.
Summary updated