SAS: Segment Anything Small for Ultrasound -- A Non-Generative Data Augmentation Technique for Robust Deep Learning in Ultrasound Imaging

arXiv:2503.05916 · eess.IV, cs.AI, cs.CV · Submitted 2026-08-24 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: Next we'll be talking about the paper "SAS: Segment Anything Small for Ultrasound -- A Non-Generative Data Augmentation Technique for Robust Deep Learning in Ultrasound Imaging".

Jane: The paper was written by Danielle L. Ferreira, Ahana Gangopadhyay, Hsi-Ming Chang, Ravi Soni and Gopal Avinash from GE HealthCare, San Ramon, California, USA.

Tom: Stay tuned as we take you through the paper and discuss its implications.

Jane: We also have Lu with us today — senior AI researcher at Tsinghua.

Tom: We also have Meng with us today — lead engineer at a mysterious AI startup.

Jane: We also have Lalam with us today — the in-house Large Language Model.

Tom: Alright, let's get started.

Summary: Tom: In the abstract of "SAS: Segment Anything Small for Ultrasound—A Non-Generative Data Augmentation Technique for Robust Deep Learning in Ultrasound Imaging," they summarize the core problem very clearly, which is that small structures are difficult to segment due to noise and variability.

Jane: They are suggesting that SAS addresses this by using a dual transformation strategy to make the training data more representative of real-world conditions.

Lu: The idea of embedding organ thumbnails into a black background is a clever way to simulate diverse scales without introducing artifacts, which is something traditional augmentation struggles with.

Meng: And then they are injecting noise into the regions of interest, which sounds like it directly addresses that variability in tissue textures we see in ultrasound scans.

Lalam: This means that if we are building diagnostic tools based on this approach, the system will be more resilient to minor fluctuations in patient anatomy or imaging quality.

Tom: The results mentioned in the summary are quite striking, showing up to a zero point three five gain in Dice score under specific conditions.

Jane: That's an impressive jump for segmentation performance; it sounds like a significant improvement over the baseline models that were used before this technique was applied.

Lu: It’s not just about the number of gains, though, Meng; we have to consider how scalable this is while looking at how they are using a foundation model.

Meng: The fact that they fine-tuned a promptable foundation model on a controlled dataset suggests that the method is designed to work with modern AI architectures.

Lalam: This ability to improve the performance of complex models like MedSAM ensures that our medical AI is moving toward state-of-the-art capabilities, which elevates the standard of care.

Improvements: Tom: Moving beyond the summary, we are looking at how SAS actually makes improvements in "SAS: Segment Anything Small for Ultrasound—A Non-Generative Data Augmentation Technique for Robust Deep Learning in Ultrasound Imaging."

Jane: The paper suggests that it improves robustness to scale variations by making the training data look more diverse, which is a major challenge in ultrasound.

Lu: They are demonstrating how this works even if you have a very limited dataset, which is huge because real-world medical data is often scarce and hard to label comprehensively.

Meng: The practical benefit here seems to be that they can generate large-scale datasets from just one thousand fifty images without needing extensive human labeling efforts.

Lalam: This efficiency means that we can deploy better AI tools faster, which contributes significantly to advancing the speed and accuracy of medical analysis across different regions.

Tom: The way they tested this by comparing a small dataset (Scenario one) against a large, diverse dataset (Scenario two) really helps us understand the impact.

Jane: It looks like SAS is able to achieve performance comparable to those larger datasets even when we start with limited data, which is quite remarkable.

Lu: The t-SNE plots they provided in the results section visually confirm that these distinct clusters are not just a theoretical concept but a tangible expansion of the feature space.

Meng: That visual evidence is important because it proves that the SAS method isn' practical; we can see how it’s generating new, unique data points in a measurable way.

Lalam: This improvement ensures that our AI models are not just good on paper, but robust enough to handle real-world variability, enhancing the overall reliability of medical imaging.

Conclusion: Tom: So, we’ve seen how "SAS: Segment Anything Small for Ultrasound—A Non-Generative Data Augmentation Technique for Robust Deep Learning in Ultrasound Imaging" tackles the problem and demonstrates significant gains.

Jane: The overall message is that even when dealing with small organs, SAS provides a powerful way to improve segmentation accuracy without falling into the trap of generating synthetic data artifacts.

Lu: I’m excited about the implications for scaling; we can have robust models trained on smaller data sets and generalize them across various unseen domains.

Meng: The engineering takeaway is that this approach is scalable, practical, and offers a way to improve model performance without increasing the computational load of massive generative models.

Lalam: It's clear that by promoting learned invariances, this work has made a huge contribution to improving how we use AI in healthcare systems.

Tom: We also saw that while SAS is great for small organs, it can't completely overcome the limitations of a very low-diversity training set like Scenario one.

Jane: That’s an important caveat; it shows that while SAS boosts generalization, it doesn' not replace the need for high-quality, diverse data when starting from a limited pool.

Lu: The iterative click prompts are another great feature to point out, showing how we can get high-quality segmentation with minimal user interaction.

Meng: It’s a highly efficient workflow that means we can integrate this into clinical systems quickly and effectively.

Lalam: We should be optimistic about the future, knowing that the foundation models will now have this specialized tool to help them achieve superior performance.

Wrap-up: Tom: As we wrap up our discussion on "SAS: Segment Anything Small for Ultrasound—A Non-Generative Data Augmentation Technique for Robust Deep Learning in Ultrasound Imaging," it's clear that SAS is a powerful and practical tool.

Jane: It’s a major step forward because, without introducing artifacts, it provides reliable improvements in segmentation across diverse anatomical structures.

Lu: I think the implications are enormous; the potential for adapting this method to other modalities is truly boundless.

Meng: It's a solid, practical solution that makes building robust AI systems much more manageable for engineers and researchers alike.

Lalam: The goal of making medical AI dependable and generalized is one we should celebrate with this kind of research, ensuring the final thought on "SAS: Segment Anything Small for Ultrasound—A Non-Generative Data Augmentation Technique for Robust Deep Learning in Ultrasound Imaging."

Danielle L. Ferreira, Ahana Gangopadhyay, Hsi-Ming Chang, Ravi Soni, Gopal Avinash

GE HealthCare, San Ramon, California, USA

eess.IV, cs.AI, cs.CV

Submitted: 2026-08-24

Updated: 2026-08-25

Importance score: 83/100

The gist: " Abstract and Problem Statement Accurate segmentation of anatomical structures in ultrasound (US) images, particularly small ones, is challenging due to "noise and variability in imaging conditions"

Key concepts

Non-Generative Data Augmentation
A method used in AI training that improves data robustness without creating synthetic artifacts. SAS utilizes this technique by applying dual transformations—like embedding organ thumbnails into a black background and injecting noise—to make training data more representative of real-world variations.
Small Structures Segmentation
The challenge of accurately segmenting small anatomical features in medical images, which is often difficult due to inherent noise and variability in the ultrasound scans. SAS specifically addresses this problem to improve diagnostic tool reliability.
Foundation Model Fine-tuning
Using a pre-trained, promptable foundation model (like MedSAM) and fine-tuning it on a controlled dataset. This suggests the method is designed to work with modern, large AI architectures, improving performance while maintaining scalability.

Terminology

Summary

"

Abstract and Problem Statement

Accurate segmentation of anatomical structures in ultrasound (US) images, particularly small ones, is challenging due to noise and variability in imaging conditions as well as size disparities. This leads to poor segmentation performance, especially for smaller organs. Conventional augmentation strategies are often insufficient because they may introduce unrealistic variations or fail to sufficiently expand the representation of small structures in training data. This necessitates a specialized augmentation technique tailored for small organ segmentation in ultrasound imaging.

Proposed Solution: Segment Anything Small (SAS)

The authors introduce Segment Anything Small (SAS), which is described as a simple yet effective scale- and texture-aware data augmentation technique designed to enhance the performance of deep learning models. SAS is a non-generative data variety approach that generates realistic and semantically consistent perturbed images, thereby avoiding the introduction of artifacts, unwanted structures, or hallucinations.

Methodology: The Dual Transformation Strategy

SAS employs a dual transformation strategy to improve data variability in terms of organ scale and tissue texture:

  1. Simulating Organ Scales (Step 1): This step simulates variations in organ size by extracting the region of interest (ROI) within the ultrasound window and rescaling it to create a thumbnail. This thumbnail is then placed on a black background matching the network input size, resulting in a scaled, zoomed-out representation of the organ. The thumbnail size is randomly selected during training, ranging from 64x64 to 256x256 pixels.

  2. Simulating Diverse Textures (Step 2): This step inject noise into the regions of interest (ROI) defined by the segmentation mask y to simulate diverse textures. The perturbation is achieved by applying one of several noise types—Speckle, Gaussian, Salt and Pepper, or Poisson—chosen at random.

Model Setup and Implementation

The authors fine-tuned a lightweight transformer-based promptable foundation model (LiteMedSAM) using SAS. They utilized an iterative click prompt technique instead of bounding boxes because bounding boxes are less effective for non-convex shapes and may inadvertently include unwanted anatomical structures. This iterative click prompting strategy simulates realistic user-generated point prompts.

The effectiveness of the models was tested under two evaluation scenarios:

  • Scenario 1: A limited single-organ dataset (1,050 images of kidney anatomy).

  • Scenario 2: A larger, more diverse multi-organ dataset (87,000 images of gynecological structures).

Experimental Results and Performance Evaluation

The results demonstrate significant improvements in segmentation performance across both small and large anatomical structures:

  • General Performance: SAS consistently improves performance across both small and large organs.

  • Quantitative Gains: In Scenario 1, SAS achieved an average improvement of 0.16 [95% CI 0.132, 0.188] for small structures and an average DSC gain of 0.084 across all test sets compared to the non-SAS baseline model (Non-SAS).

  • Small Structure Performance: When focusing on small organs, SAS achieved an average DSC improvement of 0.16 [95% CI: 0.132, 0.188] in Scenario 1 and a significant gain of up to 0.35 (in the BUSI dataset with a one-click prompt).

  • Efficiency: The iterative click prompts achieved performance comparable to bounding box prompts with just two points.

Discussion and Conclusion

The study concludes that SAS enhances model robustness and generalizability, particularly for small structures. The ability to generate large-scale datasets from a limited number of real images, while remaining free from hallucinations, diminishes the need for human labeling efforts. This makes SAS a practical and ethically sound solution for medical imaging applications. While the performance gap between SAS and non-SAS narrows when training data is increased (Scenario 2), SAS continues to provide added value, especially with minimal user interaction.

Improvements for AI systems

(Deep breath. Reviewing this bibliography indicates a strong focus on transitioning foundational computer vision models—like Segment Anything Model (SAM)—into reliable, clinically deployable medical tools. The key challenge is moving from impressive proof-of-concept results on curated datasets to robust, generalizable performance in varied, noisy clinical environments.)

Based on the convergence of these references (especially the work spanning foundation models [10]-[13], ultrasound constraints [9], efficiency [15], and evaluation rigor [30]), I propose a unified architectural improvement: The Development of a Geometrically Constrained, Low-Rank Adapted Multi-Modal Foundation Model for High-Fidelity Segmentation.

Here are the specific improvements and the resulting capabilities:

The Improvement: We must move beyond standard pixel-wise segmentation by incorporating explicit physical and geometric constraints derived from the modality (e.g., ultrasound). The system needs a dedicated module that learns the underlying biomechanical relationship between adjacent structures, rather than treating them as independent pixels.

Technical Implementation:

  1. Geometrically Constrained Augmentation Module: Utilize techniques inspired by [9] (Gudu) to generate synthetic training data that adheres to known tissue physics (e.g., maintaining smooth, curved boundaries typical of vessels or organs).

  2. Multi-Modal Fusion: Instead of running the foundation model on a single image, implement a fusion layer that accepts both the raw image input and a low-dimensional representation of the expected structural geometry (e.g., derived from Doppler data or known anatomical atlases).

What the Improved AI System Can Do:

  • Overcome Ambiguity in Ultrasound: It can reliably segment structures in highly variable modalities like ultrasound, where speckle noise and acoustic shadowing often confuse standard CNNs.

  • Enforce Anatomical Plausibility: The output segmentation map will not only be pixel-accurate but will also be physically plausible, automatically correcting segmentation errors that violate known geometric rules (e.g., a vessel cannot suddenly change diameter or cross itself).

Sources

Related papers