PSP: Harnessing Position and Shape Priors for Cross-Domain Few-Shot Medical Image Segmentation

2026-07-27 · Source: cs.CV updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Medical Image Analysis · Depth: Expert, long

Summary

The PSP (Position and Shape Priors) framework introduces a novel solution for Cross-Domain Few-Shot Medical Image Segmentation (CD-FSMIS), tackling the challenge of performance degradation caused by drastic texture discrepancies across different imaging modalities. Unlike existing methods that struggle with domain-specific intensity distributions, PSP capitalizes on the robust consistency of organ position and geometric shape across modalities. The framework integrates three core modules: a Position Coordinate Embedding (PCE) module for injecting relative spatial coordinates, a Shape Prototype Modulation (SPM) module that constructs domain-invariant structural prototypes using explicit shape priors, and a Hybrid-Prototype Prediction (HPP) module for adaptively calibrating prototypes to query feature distributions. Extensive experiments on two public medical imaging datasets demonstrated PSP's superior performance, outperforming leading methods by up to 7.05% in Dice scores on cross-sequence tasks and 3.75% on cross-modality tasks.

Key takeaway

For Machine Learning Engineers developing few-shot medical image segmentation models that struggle with cross-domain generalization, you should consider integrating explicit position and shape priors. This approach, exemplified by the PSP framework, effectively mitigates performance degradation caused by texture discrepancies between modalities. Implement modules for spatial coordinate embedding and shape prototype modulation to decouple semantic structure from domain-specific noise, significantly improving segmentation accuracy and robustness across diverse imaging scenarios.

Key insights

Harnessing invariant anatomical position and shape priors enables robust cross-domain medical image segmentation despite texture discrepancies.

Principles

• Organ position and shape are consistent across diverse medical imaging modalities.
• Decoupling semantic structure from domain-specific appearance improves generalization.
• Low-frequency shape components are more robust for cross-domain knowledge transfer.

Method

PSP uses a ResNet-50 encoder, then PCE for spatial awareness, SPM for shape-aware prototypes via geometric and spectral descriptors, and HPP for prototype calibration using a coarse-to-fine strategy.

In practice

• Use polar coordinates for position embedding in medical image features.
• Encode explicit shape priors using Signed Distance Maps and Fourier Descriptors.
• Apply dynamic residual connections to balance semantic and spatial information.

Topics

• Few-Shot Learning
• Medical Image Segmentation
• Cross-Domain Adaptation
• Position Coordinate Embedding
• Shape Prototype Modulation
• Deep Learning

Code references

• xubin471/PSP

Best for: AI Scientist, Research Scientist, Machine Learning Engineer

Related on AIssential

• VCDP: Variation-Conditioned Distributional Proxy Learning for Semi-Supervised Medical Image Segmentation — VCDP enhances semi-supervised 3D medical image segmentation by explicitly modeling both global class semantics and fine-grained intra-class variations.
• Domain Shift in Echocardiography: Interpretable Quantification and Prediction of Cross-Dataset Left Ventricular Segmentation — Echocardiographic domain shift is measurable and its impact on LV segmentation can be reduced by geometry-aware preprocessing and predicted using representation-specific metrics.
• UnDA: Unpaired Domain Alignment for Cross-Modal Knowledge Transfer in Medical Imaging — UnDA enables robust cross-modal knowledge transfer in medical imaging using unpaired data by mitigating noise and aligning features.
• ROBUST-WT: Robust Uncertainty-aware Segmentation Transform via Whitening and Training Enhancements — ROBUST-WT demonstrates that targeted training enhancements can substantially improve medical image segmentation robustness without altering core architecture.
• Active few-shot segmentation by reinforcing data selection — Reinforcement learning can optimize few-shot support sets by jointly selecting complementary examples for improved medical image segmentation.
• ConvNeXt-FD: A Fractal-Based Deep Model for Robust Biomedical Image Segmentation — ConvNeXt-FD enhances biomedical image segmentation by combining ConvNeXt with fractal-inspired boundary regularization for superior shape fidelity.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.CV updates on arXiv.org.