🤖 AI

Anatomy-Anchored Self-Supervision: Distilling Vision Foundation Models for Invariant Ultrasound Representation

This paper introduces ANAUS, an anatomy-anchored self-supervised framework that distills vision foundation models to learn invariant ultrasound representations by leveraging a prompt-engineered anatomy delineation module and a dual-policy learning strategy, thereby achieving state-of-the-art performance across multiple clinical datasets.

Original authors: Chunzheng Zhu, Yijun Wang, Jianxin Lin, Feng Wang, Hongwei Wang, Lei Zhao, Shengli Li, Kenli Li

Published 2026-05-26

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Chunzheng Zhu, Yijun Wang, Jianxin Lin, Feng Wang, Hongwei Wang, Lei Zhao, Shengli Li, Kenli Li

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to teach a robot to understand ultrasound images (the kind doctors use to look inside the body). The problem is that ultrasound images are messy. They are full of static, grainy noise, and the boundaries between organs are often fuzzy, unlike the crisp photos we see on our phones.

Most AI methods try to learn by looking at the whole image or random small squares of it. The authors of this paper argue that this is like trying to learn anatomy by looking at a blurry photo of a whole house and guessing where the kitchen is. Instead, they say we should teach the AI to focus specifically on the "rooms" (the anatomical structures) themselves.

Here is how their solution, called AnaUS, works, explained through simple analogies:

1. The "Smart Highlighter" (LP-SAM)

First, the AI needs to know what to look at. Since doctors haven't labeled millions of these images, the AI has to find the organs on its own.

The Problem: Standard AI tools are great at finding objects in clear photos (like a cat in a park) but get confused by the grainy, low-contrast nature of ultrasound.
The Solution: The team created a tool called LP-SAM. Think of this as a "Smart Highlighter."
- They took a powerful AI that already knows how to find objects in normal photos.
- They gave it a special "translator" (a learnable engine) that teaches it how to speak the language of ultrasound.
- Once trained, this tool can look at an unlabeled ultrasound image and automatically draw outlines around the heart, lungs, or thyroid without needing a human to tell it where they are. It's like giving the robot a pair of glasses that instantly makes the blurry organs stand out clearly.

2. The "Anatomy-Only" Lesson Plan

Once the AI knows where the organs are, it needs to learn what they look like. The authors use a two-part training game:

Part A: The "Same Room, Different Angle" Game (Contrastive Learning)

The Old Way: Previous methods compared random patches of an image. This meant the AI might compare a patch of a healthy heart with a patch of background noise, getting confused.
The AnaUS Way: The AI takes two slightly different views of the same ultrasound image (like looking at a statue from the left and then the right).
The Rule: It is told: "If you see the same organ (like the left ventricle of the heart) in both views, your brain should say 'These are the same!' But if you see a different organ (like the liver), you must say 'These are totally different!'"
The Result: The AI learns to ignore the grainy noise and focus strictly on the shape and texture of the specific body parts.

Part B: The "Fix the Broken Puzzle" Game (Contextual Prediction)

The Setup: The AI takes a clear picture of an organ and then "breaks" it. It doesn't just delete a random piece; it specifically damages the core of the organ using three types of "damage" that mimic real ultrasound problems:
1. Shuffling: Like moving puzzle pieces around to simulate the organ shifting under pressure.
2. Noise: Adding static to mimic the grainy interference.
3. Occlusion: Blacking out a part to mimic shadows cast by bones or air.
The Task: The AI has to look at the broken, noisy version and try to "reconstruct" what the missing center should look like.
The Result: This forces the AI to really understand the internal structure and texture of the organ, not just its surface appearance.

3. The Results

The team tested this system on six different public datasets covering lungs, breasts, thyroids, and hearts.

The Outcome: AnaUS beat every other method, including those specifically designed for ultrasound and those trained on general images.
Why it matters: It proved that when you teach an AI to focus on anatomy (the actual body parts) rather than just random image patches, it becomes much smarter and more accurate at medical tasks.

Summary

In short, the paper says: "Don't teach the AI to look at the whole messy picture. Give it a tool to find the organs first, then teach it to recognize those specific organs by comparing them from different angles and by practicing how to fix them when they get damaged." This approach creates a much stronger foundation for medical AI without needing millions of human-labeled examples.

Technical Summary: Anatomy-Anchored Self-Supervision for Ultrasound Representation Learning

1. Problem Statement

Ultrasound (US) imaging is a clinical cornerstone due to its portability and non-ionizing nature, yet deep learning applications are hindered by the scarcity of annotated data. While Self-Supervised Learning (SSL) has emerged as a solution, existing methods for US images suffer from two primary limitations:

Frame-level approaches: These leverage temporal coherence in video but treat frames holistically, often encoding background artifacts and speckle noise rather than diagnostically relevant anatomy.
Region-level approaches: Borrowed from natural images, these rely on generic segmentation (e.g., K-means, image-computable proposals) that fail to isolate meaningful anatomical structures in US images, which are characterized by low-contrast boundaries and speckle corruption.

Neither paradigm aligns with clinical reasoning, which is inherently anatomy-centric. The authors argue that representation learning must prioritize anatomical structures over stochastic image patches to achieve clinically aligned representations.

2. Methodology: The AnaUS Framework

The proposed AnaUS (Anatomy-Anchored UltraSound Self-Supervision) framework shifts representation learning from generic visual regions to clinically meaningful anatomical structures. The pipeline consists of three core components:

A. Anatomy Discovery via LP-SAM (Latent-Prompted SAM)
To address the domain gap between natural images and ultrasound, the authors introduce LP-SAM, a module that enables annotation-free anatomical delineation.

Stage 1 (Adapter-based Domain Specialization): Learnable adapters are inserted into the ViT blocks of the pre-trained Segment Anything Model (SAM) image encoder. These adapters are fine-tuned on existing public US image–mask pairs (DDTI, TG3K, CAMUS) using a combined Binary Cross-Entropy and Dice loss to handle foreground-background imbalance.
Stage 2 (Latent Prompt Engine - LPE): To enable fully automated, prompt-free mask generation, a Cross-Perception Attention (CPA) module fuses global ViT features and local CNN features. A learnable query token is iteratively refined through bidirectional cross-attention to generate a latent prompt embedding. This embedding guides SAM's mask decoder to produce anatomical masks without manual prompts.
Outcome: LP-SAM generates a set of anatomical masks $M(x) = \{m_k\}$ for unlabeled US images, serving as the foundation for the SSL process.

B. View-Invariant Anatomical Contrastive Matching
Instead of comparing entire images or random crops, AnaUS performs contrastive learning at the level of identified anatomical structures.

Mask Alignment: Two augmented views ( $v_1, v_2$ ) of an image are generated, and their corresponding anatomical masks are aligned.
Multi-scale Mask Pooling: Features are extracted from a shared backbone at multiple spatial scales. For each anatomical mask, a pooled hidden vector is computed by averaging features within the region of interest, effectively eliminating background contamination.
Contrastive Objective: An asymmetric online–target architecture (similar to BYOL) is employed. The loss function encourages feature invariance for the same anatomical structure across different views while maximizing separation (discriminability) between distinct anatomical structures within the same image. This is computed across multiple scales to capture both coarse structural and fine textural details.

C. Contextual Prediction over Anatomical Regions
To capture structural context beyond surface appearance, a reconstruction task is introduced.

Targeted Corruption: Unlike random masking, this task targets the "core" anatomical region (the overlapping crop between augmented views).
Damage Function: A specialized damage function $D(I; \alpha, \beta, \gamma)$ $D (I; α, β, γ)$ simulates US-specific artifacts:
- Pixel Shuffling: Simulates tissue displacement under probe pressure.
- Gaussian Noise: Mimics speckle interference.
- Occlusion: Emulates acoustic shadowing.
Objective: The encoder is compelled to reconstruct the corrupted core region from the intact view, forcing the model to internalize local textural intricacies and global structural coherence.

The total pre-training loss combines the anatomy-level contrastive loss ( $L_{ana}$ ) and the contextual prediction loss ( $L_{ctx}$ ).

3. Key Contributions

The authors summarize their contributions as follows:

AnaUS Framework: The first anatomy-level self-supervised pre-training framework explicitly designed for ultrasound, shifting the contrastive unit from generic regions to clinically relevant structures.
LP-SAM Module: Introduction of a learnable latent prompt engine that enables fully automated, expert-free structural discovery in ultrasound imagery through targeted latent-space adaptation, bridging the natural-to-medical domain gap.
Performance Benchmark: Demonstration of consistent state-of-the-art performance across diverse US datasets and tasks, establishing a new benchmark for ultrasound self-supervised pre-training.

4. Experimental Results

The framework was evaluated on six public datasets covering lung, breast, thyroid, and cardiac imaging.

Classification: On the POCUS dataset, AnaUS achieved 96.2% accuracy, surpassing the strongest US-specific baseline (IVPP++) by +3.0%. On the BUSI dataset, it reached 87.1% accuracy, outperforming IVPP++ by +3.6%.
Segmentation: On the UDIAT-B benchmark, AnaUS achieved 85.9% Dice and 92.7% PPV. It consistently outperformed general SSL methods, object-centric methods, and previous US-specific pre-training approaches.
Ablation Studies:
- Replacing LP-SAM with vanilla SAM resulted in a sharp performance drop (−15.7% Dice), confirming the necessity of domain adaptation.
- Removing the symmetric contrast loss reduced Dice by 8.3%.
- Removing the contextual prediction task reduced Dice by 1.0%.
- Replacing generic mask sources (e.g., K-means) with LP-SAM masks significantly boosted object-centric baselines, confirming that anatomically precise masks are a decisive factor for pre-training quality.
Qualitative Analysis: Visual results showed AnaUS maintained sharp boundary delineation for challenging structures (e.g., thyroid nodules, myocardial edges) where other methods exhibited over-smoothing or coarse localization. Zero-shot generalization tests confirmed LP-SAM's ability to discover unseen anatomical structures (e.g., "bat sign" in lung, LA/MYO in cardiac).

5. Significance and Claims

The paper claims that AnaUS underscores the pivotal role of anatomically meaningful priors in guiding representation learning for medical imagery. By shifting the focus from stochastic image patches to clinically relevant structures, the framework aligns self-supervised learning with clinical reasoning. The authors assert that their work establishes a principled and clinically aligned foundation for future research in medical self-supervised learning, offering a computationally efficient solution suitable for clinical deployment.