Domain Generalized Medical Landmark Detection via Robust Boundary-Aware Pre-Training

Haifan Gong, Yu Lu, Xiang Wan, Haofeng LiAAAI 2025 · DOI

Introduction

Medical landmark detection is a critical step in clinical workflows, enabling precise anatomical measurement and surgical planning. However, models often suffer from a significant "domain gap" when deployed across different hospitals. Variations in MRI/CT scanners, imaging protocols, and patient demographics lead to performance degradation on unseen data. While Unsupervised Domain Adaptation (UDA) offers a solution, it requires access to target domain data during training—a luxury often prohibited by strict data privacy regulations.

This paper addresses Domain Generalization (DG), aiming to train a robust model on a single source dataset that generalizes to any unseen domain. The authors propose BaPDG (Boundary-aware Pre-training for Domain Generalization), a framework designed to overcome low contrast, high noise, and the inherent task disparity between organ segmentation and landmark localization.

The BaPDG Framework

The core innovation of BaPDG is a two-stage training strategy that leverages anatomical structures to guide landmark detection.

Overview of the BaPDG framework, illustrating the two-stage process: boundary-aware pre-training using distance maps and fine-tuning for landmark detection.

1. Boundary-Aware Pre-Training (BaP)

Most landmarks are located on anatomical boundaries. BaPDG exploits this by "priming" the network with existing segmentation datasets. Instead of binary masks, the model is trained to predict Boundary Distance Maps. These maps represent the distance of each pixel to the nearest organ boundary, forcing the model to learn the underlying geometric structure of the anatomy before specializing in specific landmark points.

2. Robust Mixed Loss Function

Standard Mean Squared Error (MSE) loss can be overly sensitive to large initial errors, potentially leading to poor generalization. The authors introduce a time-adaptive mixed loss: $$L = \frac{e}{E} \text{MSE} + \frac{E-e}{E} \text{logcosh}$$ where $e$ is the current epoch and $E$ is the total number of epochs. The Log-cosh loss provides robustness early in training by behaving like Mean Absolute Error (MAE) for large residuals, while the MSE loss ensures high precision as the model converges.

Theoretical Grounding

The paper provides a formal proof using Rademacher complexity and Lipschitz continuity to demonstrate that the log-cosh loss provides a tighter generalization bound than MSE. By reducing the "curvature" (Hessian) of the loss surface, the model is less likely to overfit to noise in the source domain, thereby improving its performance on unseen data.

Experimental Results

The authors evaluated BaPDG on a new multi-center benchmark, DGMLD, which includes fetal MRI scans from three different centers (Atlas, LFC, and FeTA).

Examples of anatomical landmarks and boundaries in fetal brain MRI, showing the structural context used for robust detection.

Key findings include:

  • Superior Generalization: On the external FeTA dataset, BaPDG achieved a Mean Radial Error (MRE) of 2.75mm using a UNet backbone, a significant improvement over the baseline's 6.27mm.
  • High Precision: The Successful Detection Rate (SDR) curves indicate that BaPDG consistently identifies more landmarks within tight error tolerances (e.g., 2mm) compared to state-of-the-art methods like AugDG and EdgeDG.
  • Cross-Modality Validation: The method also showed strong results on 2D X-ray cephalometric data, proving its versatility beyond 3D MRI.

Conclusion

BaPDG offers a robust solution for domain-generalized medical landmark detection by bridging the gap between segmentation and localization. By repurposing widely available segmentation masks for boundary-aware pre-training and employing a theoretically grounded loss function, the framework achieves state-of-the-art performance across diverse imaging centers without requiring target-domain data.

Summary updated

BibTeX
@inproceedings{gong2025domain,
  author = {Haifan Gong and Yu Lu and Xiang Wan and Haofeng Li},
  title = {{Domain Generalized Medical Landmark Detection via Robust Boundary-Aware Pre-Training}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2025},
  doi = {10.1609/aaai.v39i3.32323},
}
Figure