Healthcare
Deep Learning for Medical Ultrasound Image Segmentation: A Systematic Review of the Current Research
This systematic review analyzes 296 recent scientific articles on Deep Learning (DL) for medical ultrasound image segmentation, following PRISMA 2020 guidelines. It identifies breast tumors, organs, and cardiovascular structures as the most common targets. While Convolutional Neural Networks (CNNs), especially U-shaped architectures, remain popular, newer models like vision transformers (ViTs), CNN/ViT hybrids, and Segment Anything Models (SAMs) are gaining ground. The review highlights that newer models often utilize significantly more data, yet no clear association between method type and evaluation metrics is observed across studies. Common limitations include a lack of information on computational requirements and issues in model performance evaluation. The field is rapidly evolving due to increased ultrasound use, new public datasets, and methodological advancements.
Executive Impact
This research provides a critical overview for healthcare enterprises looking to integrate AI into medical imaging. By understanding current trends, popular models, and identified gaps, organizations can strategically invest in DL solutions for ultrasound segmentation, optimizing diagnostics and operational efficiency. The analysis of model performance and data requirements helps inform robust AI strategy development.
296
Articles Analyzed
3
Primary Targets
3
Newer Models Adopted
Deep Analysis & Enterprise Applications
Select a topic to dive deeper, then explore the specific findings from the research, rebuilt as interactive, enterprise-focused modules.
The review categorizes Deep Learning methods into CNNs, ViTs, CNN/ViT hybrids, Segment Anything Models (SAMs), and Generative Models. CNNs, particularly U-Net variants, maintain popularity, especially in oncology and obstetrics. ViTs and hybrid models, though newer, are establishing their presence. SAMs are noted for their rapid adoption despite recent release. Generative models are less common.
59.1
Studies used CNNs (%)
7.4
Studies used ViTs (%)
| Method Type | Popularity | Characteristics | Typical Data Needs |
|---|---|---|---|
| CNNs (U-Net variants) |
|
|
|
| ViTs |
|
|
|
| CNN/ViT Hybrids |
|
|
|
| SAMs |
|
|
|
The review identifies breast tumors, organ segmentation, and cardiology as the top application areas for DL-based ultrasound segmentation. Other significant areas include angiology, obstetrics, orthopedics, thyroid nodules, oncology, and gynecology. Breast tumor segmentation is particularly dominant, often serving as a testbed for new methods.
25.0
Breast Tumor Segmentation (%)
10.8
Organ Segmentation (%)
Application Focus Progression
Breast Tumors (25%)
→
Organ Segmentation (10.8%)
→
Cardiology (9.8%)
→
Angiology (9.8%)
→
Obstetrics (9.5%)
→
Orthopedics (8.8%)
→
Thyroid Nodules (8.8%)
→
Oncology (7.1%)
→
Gynecology (5.1%)
The study reveals a trend towards using more data for newer models, with public datasets being more common in breast tumor segmentation. Despite this, no significant association was found between method type or data size and reported Dice scores, suggesting other confounding factors or limitations in current evaluation practices.
38.5
Studies used Public Datasets (%)
4932+
Mean Data Instances
Understanding Model Performance & Data Impact
The review observed that studies utilizing public datasets had significantly more data instances than those with private data (p-value, 8.8e-5). Furthermore, ViTs and CNN/ViT hybrids also included significantly more data instances than CNNs (p-value, 0.031). However, surprisingly, no statistically significant association was found between the number of data instances and the reported mean Dice scores (Spearman’s correlation coefficient of -0.00991 with a p-value of 0.879). This suggests that while newer models are given more data, the direct impact on reported Dice scores across diverse studies is not clear, potentially due to confounding factors, differences in specific segmentation tasks, or limitations in how results are aggregated and compared. This highlights a crucial gap for enterprises: simply having more data or using a 'newer' model does not guarantee superior reported performance across all contexts, and careful validation tailored to specific clinical applications is essential.
Quantify Your AI Impact
Estimate the potential annual cost savings and efficiency gains your enterprise could achieve by automating image segmentation with AI.
Potential Annual Savings
$-
Annual Hours Reclaimed
-
AI Implementation Roadmap for Ultrasound Segmentation
A strategic phased approach to integrating Deep Learning solutions for medical ultrasound image segmentation into your enterprise.
Phase 1: Assessment & Strategy (2-4 Weeks)
Conduct a thorough needs assessment, define clear segmentation objectives, evaluate existing infrastructure, and develop a comprehensive AI strategy aligned with clinical goals.
Phase 2: Data Curation & Annotation (4-12 Weeks)
Identify, collect, and meticulously annotate ultrasound datasets. Prioritize high-quality, diverse data that represents the target anatomy and pathologies. Establish data governance.
Phase 3: Model Selection & Customization (6-16 Weeks)
Select or customize a Deep Learning model (e.g., U-Net, ViT hybrid) based on your data characteristics and performance requirements. Address computational resources and model interpretability.
Phase 4: Training & Validation (8-20 Weeks)
Train the selected model using curated datasets. Implement rigorous validation protocols, including independent testing, to ensure accuracy, robustness, and generalizability across diverse patient populations.
Phase 5: Integration & Deployment (4-8 Weeks)
Integrate the validated AI model into your existing PACS or clinical workflow. Develop monitoring tools for ongoing performance, user feedback mechanisms, and secure deployment strategies.
Phase 6: Post-Deployment Optimization & Scaling (Ongoing)
Continuously monitor model performance, collect new data for retraining, and iterate on improvements. Explore scaling the solution to other imaging modalities or clinical applications.
Ready to explore how Deep Learning can transform your medical imaging workflows? Book a personalized consultation with our AI specialists.
Schedule Your Strategy SessionReady to Get Started?
Book Your Free Consultation.
Let's Discuss Your AI Strategy!
Lets Discuss Your Needs
Select a Date
Sun
Mon
Tue
Wed
Thu
Fri
Sat