Enterprise AI Analysis
SFP-TFF: A Spatial-Frequency-Phase Tensor Fusion Framework for Semi-Supervised Medical Image Segmentation
This paper introduces SFP-TFF, a novel Spatial-Frequency-Phase Tensor Fusion Framework for semi-supervised medical image segmentation. It addresses the critical challenge of accurate medical image segmentation with limited annotations by fusing spatial, frequency, and phase domain features. The framework uses a dual-branch architecture, a tensor fusion module, and a semi-supervised learning strategy to enhance feature complementarity and leverage unlabeled data. Extensive experiments confirm superior performance in low-label and challenging imaging conditions, leading to more robust and generalized medical image segmentation.
Authors: SONGHE YUAN, LAURENCE T. YANG, DEBIN LIU, YUCHENG GAO
Published: 09 March 2026 | Accepted: 26 January 2026 | Total Citations: 0 | Total Downloads: 0
Executive Impact at a Glance
SFP-TFF's innovative fusion of spatial, frequency, and phase information delivers transformative benefits for medical AI applications, significantly enhancing accuracy and efficiency where it matters most.
0%
AI-Driven Efficiency Gain
0x
ROI Potential (First Year)
0%
Reduced Annotation Effort
0%
Error Rate Reduction
Deep Analysis & Enterprise Applications
Select a topic to dive deeper, then explore the specific findings from the research, rebuilt as interactive, enterprise-focused modules.
Core Innovation: SFP-TFF Framework
The SFP-TFF framework introduces a novel approach to semi-supervised medical image segmentation by integrating spatial, frequency, and phase information. This multi-domain fusion addresses the limitations of relying solely on spatial features, which often struggle with noise and ambiguous boundaries in medical images. By leveraging frequency-phase information, the model achieves more robust and discriminative representations, particularly crucial in scenarios with limited labeled data.
8.8%
DSC Improvement (Over baseline methods in low-label settings (ACDC).)
60.5%
Boundary Adherence (HD95 Reduction (ACDC) with 10% labels.)
Frequency-Phase Decomposition Module
This module extends the classical Fourier Transform to decompose encoder features into high-frequency, low-frequency, and phase components. High-frequency captures fine-grained details and edges, low-frequency retains anatomical structure, and phase encodes relative positional relationships. This disentanglement helps create noise-resilient and structure-preserving representations, vital for complex medical imagery.
High
Noise Resilience (Improved robustness against imaging noise.)
Excellent
Structural Preservation (Phase information enhances boundary localization.)
Tucker Fusion with Channel Attention
A channel-aware Tucker tensor fusion module integrates the multi-domain features. Unlike traditional methods, it adaptively reweights spatial, spectral, and phase contributions, modeling high-order cross-domain interactions with controllable complexity. This prevents uniform domain contribution and emphasizes informative cues critical for boundary delineation, leading to more discriminative segmentation representations.
Adaptive
Multi-Domain Integration (Channel-aware reweighting for optimal fusion.)
Low-Rank Tucker
Complexity Control (Efficiently models high-order interactions.)
Frequency-Phase Guided Contrastive Learning
The Frequency-Phase Guided Contrastive Learning (FPCL) module constructs semantically meaningful contrastive pairs based on frequency amplitude and phase orientation, not just pixel-wise annotations. This strategy identifies informative and structurally relevant anchors using spectral saliency and phase consistency criteria, improving segmentation performance in the presence of annotation noise and boundary ambiguity by enforcing structural alignment.
High
Annotation Efficiency (Leverages unlabeled data more effectively.)
Enhanced
Semantic Alignment (Guided by frequency-phase priors.)
Multi-Level Loss Function Design
The training framework employs a comprehensive multi-level loss function combining supervised loss (Dice + BCE), multi-domain consistency loss (frequency, phase, mask consistency), frequency-phase alignment loss, and auxiliary deep supervision. This collaborative optimization guides the model towards learning robust, anatomically consistent features and improves training stability under low-label regimes.
Robust
Training Stability (Through multi-faceted regularization.)
Maximized
Segmentation Accuracy (Comprehensive loss function design.)
Performance & Validation
Extensive experiments on benchmark datasets (ACDC, UDIAT, LUNA16) demonstrate SFP-TFF's superior performance over state-of-the-art methods, particularly in low-label settings. The framework consistently achieves higher DSC scores and lower HD95/ASD, indicating better segmentation accuracy and boundary adherence across diverse medical imaging modalities (MRI, ultrasound, CT) and anatomical structures.
0.896
DSC Score (10% Labels, ACDC dataset (Our method).)
60.5%
HD95 Reduction (10% Labels, ACDC dataset (from 10.79 to 4.26).)
15-20%
AI-Driven Efficiency Gain in Medical Image Analysis
SFP-TFF consistently outperforms existing state-of-the-art methods, particularly in low-label settings and challenging imaging conditions, leading to more robust and accurate segmentations with reduced manual effort.
Enterprise Process Flow
Input Medical Image (Labeled/Unlabeled)
→
Student/Teacher Network Processing
→
Frequency-Phase Decomposition
→
Spatial-Frequency-Phase Feature Aggregation
→
Channel-Aware Tucker Fusion
→
Frequency-Phase Guided Contrastive Learning
→
Multi-Level Loss Optimization
→
Accurate Segmentation Mask Output
| Feature | Traditional SSL Methods | SFP-TFF (Our Method) |
|---|---|---|
| Feature Domains Utilized |
|
|
| Handling Low-Contrast/Noise |
|
|
| Annotation Efficiency |
|
|
| Feature Integration |
|
|
| Boundary Delineation |
|
|
SFP-TFF in Cardiac MRI Segmentation (ACDC Dataset)
In complex cardiac MRI segmentation (ACDC dataset), SFP-TFF demonstrates superior performance in delineating structures like the right ventricle (RV), myocardium (MYO), and left ventricle (LV). Compared to baselines, our method avoids under-segmentation and contour fragmentation, achieving a 60.5% reduction in HD95 with only 10% labeled data. This is crucial for downstream functional quantification (e.g., ejection fraction calculation) and clinical reporting, where precise anatomical measurements are paramount. The integration of frequency-phase information proves particularly effective in handling the inherent low contrast and subtle boundaries characteristic of cardiac MRI.
Key Results:
- Achieved DSC of 0.896 with only 10% labeled data (compared to 0.823 for TCSM baseline).
- Reduced Hausdorff Distance (HD95) from 10.79 to 4.26, indicating significantly improved boundary adherence.
- Enabled robust segmentation under highly challenging imaging conditions with limited supervision.
Calculate Your Potential ROI
See how integrating SFP-TFF can translate into significant annual savings and reclaimed human hours for your organization.
Your Implementation Roadmap
A structured approach to integrating SFP-TFF into your existing infrastructure for seamless transition and maximum impact.
Phase 01: Initial Assessment & Strategy
Evaluate your current medical imaging workflows, data infrastructure, and specific segmentation challenges. Define key performance indicators (KPIs) and tailor an SFP-TFF integration strategy.
Phase 02: Data Preparation & Model Training
Prepare your labeled and unlabeled medical image datasets. Deploy and fine-tune SFP-TFF with specialized training, leveraging its semi-supervised capabilities for efficient model adaptation.
Phase 03: Pilot Deployment & Validation
Conduct a pilot deployment within a controlled environment, validating SFP-TFF's performance against defined KPIs using real-world data and expert feedback.
Phase 04: Full-Scale Integration & Optimization
Integrate SFP-TFF into your production systems, optimizing for scalability, efficiency, and ongoing performance. Establish continuous monitoring and maintenance protocols.
Ready to Transform Your Enterprise?
Unlock the full potential of advanced medical image segmentation. Our experts are ready to guide you through a personalized strategy session.
Ready 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