ProPL: Universal Semi-Supervised Ultrasound Image Segmentation via Prompt-Guided Pseudo-Labeling

Conference: AAAI 2026 arXiv: 2511.15057 Code: https://github.com/WUTCM-Lab/ProPL Area: Medical Imaging / Ultrasound Segmentation Keywords: universal segmentation, semi-supervised learning, pseudo-labeling, prompt guidance, ultrasound imaging

TL;DR

ProPL is proposed as a framework that, for the first time, achieves universal semi-supervised ultrasound image segmentation via a shared visual encoder, prompt-guided dual decoders, and uncertainty-driven pseudo-label calibration. With only 1/16 labeled data across 5 organs and 8 tasks, it surpasses fully supervised methods by 5.18% mDice.

Background & Motivation

Background: Ultrasound image segmentation is critical for computer-aided diagnosis, yet existing methods are typically designed for specific organs or tasks, exhibiting poor generalizability.

Limitations of Prior Work: - Fully supervised methods require large amounts of annotated data, which is particularly costly for ultrasound images due to speckle noise, acoustic shadowing, and tissue artifacts that blur boundaries. - Semi-supervised methods reduce annotation requirements but remain confined to single tasks. - Universal segmentation frameworks (e.g., DoDNet, UniSeg) support only fully supervised settings and are thus constrained by annotation availability.

Core Problem: How to construct a universal ultrasound segmentation framework capable of handling multiple organs and tasks simultaneously while requiring only minimal annotations?

Key Insight: Combining prompt learning for task adaptation with dual-decoder mutual learning to generate reliable pseudo-labels.

Method

Overall Architecture

Input ultrasound images → shared ConvNeXt-Tiny encoder → dual decoders (standard decoder \(\mathcal{G}_{sd}\) + prompt decoder \(\mathcal{G}_{pd}\)) → mutual pseudo-label supervision between decoders. Task prompts are encoded via BERT and injected into the prompt decoder.

Key Designs

1. Prompting-upon-Decoding (PuD):

   • Function: Injects task-specific textual prompts into the decoding process.
   • Mechanism: Task descriptions are encoded via BERT to obtain \(\bm{t}\), aligned in dimensionality through 1D convolution and linear projection, and then injected into decoding features via multi-head cross-attention: \(\bm{h}_k = \bm{z}_k' + \alpha \cdot \text{MHCA}(Q=\bm{z}_k', K=\bm{\tau}, V=\bm{\tau})\)
   • Design Motivation: Unlike one-hot encodings or learnable prompts, textual prompts carry richer semantics and generalize to unseen tasks; the learnable scalar \(\alpha\) controls the influence of the prompt.

2. Uncertainty-driven Pseudo-Label Calibration (UPLC):

   • Function: Filters and calibrates pseudo-labels based on prediction uncertainty.
   • Mechanism: Each decoder independently generates predictions; regions of disagreement are treated as high-uncertainty, and only high-confidence regions are used for mutual pseudo-label supervision.
   • Design Motivation: Raw pseudo-labels introduce noise; the divergence between the two decoders serves as an uncertainty signal for filtering unreliable predictions.

3. Universal Ultrasound Dataset:

   • 6,400 images spanning 5 organs (breast, fetal, cardiac, ovary, thyroid) across 8 segmentation tasks.
   • Labeled data partitions: 1/16, 1/8, and 1/4 settings.

Key Experimental Results

Main Results (1/16 Labeled Data)

Method Type	Method	mDice%	mIoU%
Single-task supervised	U-Net	75.17	64.76
Single-task semi-supervised	UniMatch	79.38	69.66
Universal supervised	DoDNet	62.99	50.04
Universal supervised	CLIP-UM	63.70	51.27
Universal semi-supervised	ProPL	80.35	70.63

Different Labeling Ratios

Labeling Ratio	ProPL mDice	Gain over Runner-up
1/16	80.35%	+0.97%
1/8	82.56%	+2.2%
1/4	83.70%	+1.32%

Ablation Study (1/16)

Configuration	mDice	mIoU
w/o prompt (PuD)	60.76	52.57
w/o UPLC	77.85	67.23
Full ProPL	80.35	70.63

Key Findings

• Removing task prompts causes a 19.59% drop in mDice (80.35→60.76), underscoring the critical role of prompts in universal models.
• UPLC contributes a 2.5% mDice improvement, confirming the effectiveness of uncertainty-based pseudo-label filtering.
• ProPL requires only 712 MB of GPU memory and dominates the performance–efficiency Pareto frontier over all baselines.
• Universal supervised methods (DoDNet, CLIP-UM) perform poorly on ultrasound data (~63% mDice), indicating that universal models benefit substantially from semi-supervised augmentation.

Highlights & Insights

• First formulation of "universal semi-supervised ultrasound segmentation": integrates multi-organ multi-task generality with the practical constraint of limited annotations.
• Textual prompts vs. one-hot/learnable prompts: textual prompts add only 18 s/epoch overhead but provide substantially richer semantics; their removal causes model collapse.
• Dual-decoder mutual learning: high-confidence predictions from one decoder serve as pseudo-labels for the other, with inter-decoder disagreement used to estimate uncertainty.

Limitations & Future Work

• The dataset covers only 2D ultrasound images and does not extend to 3D volumetric ultrasound.
• Prompt templates require manual design; automated prompt generation may yield further improvements.
• The threshold in UPLC requires manual tuning; adaptive threshold strategies warrant further exploration.
• Cross-modality generalization (e.g., whether the framework transfers to CT/MRI) remains unvalidated.

Related Work & Insights

• vs. UniMatch (semi-supervised): UniMatch is the single-task semi-supervised state of the art; ProPL surpasses it by 0.97% in the universal setting.
• vs. DoDNet/UniSeg (universal supervised): These methods perform poorly on ultrasound data; ProPL compensates for annotation scarcity through semi-supervised learning.
• vs. SAM-based methods: SAM-based approaches require additional interactive prompts (points, scribbles), whereas ProPL requires only textual task descriptions.

Rating

• Novelty: ⭐⭐⭐⭐ First definition of universal semi-supervised ultrasound segmentation with a well-motivated framework design.
• Experimental Thoroughness: ⭐⭐⭐⭐⭐ Eight tasks, multiple labeling ratios, detailed ablations, and efficiency analysis.
• Writing Quality: ⭐⭐⭐⭐ Problem formulation is clear and contributions are well-articulated.
• Value: ⭐⭐⭐⭐ Both the dataset and the framework offer practical contributions to clinical ultrasound segmentation.

Related Papers

• [AAAI 2026] Bidirectional Channel-selective Semantic Interaction for Semi-Supervised Medical Segmentation
• [CVPR 2026] A Semi-Supervised Framework for Breast Ultrasound Segmentation with Training-Free Pseudo-Label Generation and Label Refinement
• [AAAI 2026] DeNAS-ViT: Data Efficient NAS-Optimized Vision Transformer for Ultrasound Image Segmentation
• [AAAI 2026] DualFete: Revisiting Teacher-Student Interactions from a Feedback Perspective for Semi-supervised Medical Image Segmentation
• [ACL 2026] Semi-Supervised Diseased Detection from Speech Dialogues with Multi-Level Data Modeling