Research map / Detection and segmentation / Open-set and few-shot detection
Domain Generalized Medical Landmark Detection via Robust Boundary-Aware Pre-Training
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes Boundary-aware Pre-training (BaP) that pre-trains a landmark detector on boundary distance maps from public single-center segmentation data (Atlas) with a log-cosh+MSE mixed loss, and introduces a DG medical landmark benchmark (DGMLD) for fetal cerebellum landmarks.
Problem
Medical landmark detection models degrade across imaging centers (domain shift), and existing work addresses single-center or unsupervised domain adaptation rather than domain generalization where only single-source labeled data is available.
Approach
Pre-train the network on boundary distance maps derived from organ segmentation masks (since landmarks lie on structure boundaries), using a loss that linearly shifts from log-cosh (robust early) to MSE (later), then fine-tune on Gaussian heatmaps for landmark detection.
Key results
- On the DGMLD benchmark with a UNet backbone, BaPDG reaches average MRE 2.67 on LFC (vs. 2.88 baseline) and 2.75 on FeTA (vs. 6.27 baseline).
- With ViTPose on FeTA, BaPDG achieves average MRE 5.38 versus 8.16 for the baseline and 7.68 for SegDG.
- On 2D cephalometric landmarks, the method reaches MRE 2.66 on out-of-domain PKU versus 3.02 for the baseline.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use BaPDG instead of AugDG | UNet · LFC | MRE: 2.87±0.23 -> 2.67±0.09 | Helped |
| Use BaPDG instead of EdgeDG | ViTPose · LFC | MRE: 3.62±0.31 -> 3.21±0.12 | Helped |
| Use BaPDG instead of SegDG | UNet · FeTA | MRE: 3.15±0.38 -> 2.75±0.48 | Helped |
| Use BaPDG instead of AugDG | ViTPose · FeTA | MRE: 6.28±1.00 -> 5.38±0.75 | Helped |
| Use the proposed method instead of EdgeDG | 2D cephalometric landmark detection · ISBI-15 (in domain) | MRE: 1.98 -> 1.95 | Helped |
| Use the proposed method instead of Baseline | 2D cephalometric landmark detection · PKU (out of domain) | SDR 3mm: 74.09 -> 74.25 | Helped |
| Replace direct segmentation-mask pre-training with boundary distance-map pre-training using MSE (ablation) | BaPDG pre-training · LFC validation | MRE: 2.14±0.05 -> 2.05±0.07 | Helped |
| Add log-cosh loss to MSE during boundary-distance-map pre-training (ablation) | BaPDG pre-training · LFC validation | MRE: 2.05±0.07 -> 1.97±0.14 | Helped |
Surprising findings
- Direct supervision with segmentation masks (SegDG) gives little or no gain over baseline, and EdgeDG sometimes degrades performance on FeTA (UNet 6.08 vs 6.27 average, but worse on several landmarks).
Limitations
- Pre-training relies on Atlas with only 40 segmentation-labeled cases and a single source center.
- Improvements are not uniform across all landmarks and backbones (e.g., UNet on LFC ADV2 and FeTA variance is high).
- Theoretical generalization analysis is loose (Hoeffding-based bound, Lipschitz argument for MSE is data-dependent).
Experimental setup
Datasets: Atlas LFC FeTA ISBI-15 PKU
Benchmarks: DGMLD (LFC, FeTA) ISBI-15 PKU
Metrics: MRE SDR
Baselines: Baseline AugDG EdgeDG SegDG EdgeDG (2D cephalometric)
Builds on: Gaussian heatmap landmark regression (Li et al. 2023) AugDG (Avisdris et al. 2022) EdgeDG (Gu et al. 2024) UNet
Closest papers by meaning
- Why Invariance is Not Enough for Biomedical Domain Generalization and How to Fix It (NeurIPS 2026 · similarity 0.88)
- Learning Generalized Medical Image Segmentation from Decoupled Feature Queries (AAAI 2024 · similarity 0.87)
- Training Like a Medical Resident: Context-Prior Learning Toward Universal Medical Image Segmentation (CVPR 2024 · similarity 0.86)
- Multi-view Evidential Learning-based Medical Image Segmentation (AAAI 2025 · similarity 0.85)
- Modality-agnostic Domain Generalizable Medical Image Segmentation by Multi-Frequency in Multi-Scale Attention (CVPR 2024 · similarity 0.85)
- Generative Medical Segmentation (AAAI 2025 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.