Enabling Real-Time Point-of-Care Ultrasound Segmentation: A GPU-Free Deployment in Resource-Limited Settings

Technical Summary: Enabling Real-Time Point-of-Care Ultrasound Segmentation via GPU-Free Deployment

1. Problem Statement

Ultrasound imaging is the most widely adopted medical modality globally due to its low cost, portability, and lack of ionizing radiation. However, a structural paradox exists in its integration with Artificial Intelligence (AI): while the imaging devices are affordable, the deployment of AI-driven diagnostic tools relies heavily on GPU-accelerated, heavy-weight models (e.g., TransUNet, nnUNet) with tens to hundreds of millions of parameters. This creates a "hardware accessible, intelligence inaccessible" scenario, particularly in resource-limited settings (LMICs), primary care clinics, and mobile health scenarios where GPU infrastructure is absent. Furthermore, the operator-dependent nature of ultrasound exacerbates diagnostic inconsistencies, and the high cost of AI deployment prevents the democratization of advanced diagnostics in the very regions that need them most.

2. Methodology

The authors present UltraSeg, an ultra-lightweight segmentation architecture originally developed for colonoscopic polyp segmentation, systematically adapted for Point-of-Care Ultrasound (POCUS). The methodology focuses on architectural efficiency rather than post-hoc compression.

• Architecture Design: UltraSeg utilizes an Encoder-Decoder topology adhering to "parametric minimalism." Key components include:
  • Multi-scale Context Aggregation: An Enhanced Dilated Block (EDB) at the third encoding layer uses parallel dilated convolutions to expand the receptive field with negligible parameter cost.
  • Cross-layer Feature Fusion: Employs Attention-Guided Fusion (AGF) and Simple Spatial Attention (SSA) to merge shallow details with deep semantics.
  • Global Context Modeling: Predictive Gated Fusion (PGF) and Group Shuffle Attention (GSA) are integrated into the upsampling stages to enhance edge perception and reduce parameters.
  • Dual Deep Supervision: Joint optimization of region and boundary losses (Dice and Boundary Loss) to improve contour delineation in low-contrast ultrasound images.
• Domain Adaptation for Ultrasound:
  • Receptive Field Adjustment: To accommodate the larger field-of-view and deeper tissue location of ultrasound lesions, the EDB dilation rates were modified from the original colonoscopy configuration [1, 2, 3] to [2, 4, 6]. This expands the effective receptive field from ~50 to ~100 pixels without adding parameters.
  • Dual-Class Adaptation: For obstetric tasks requiring simultaneous segmentation of two structures (e.g., fetal head and pubic symphysis), the architecture was modified to support multi-class outputs via dynamic channel processing and hierarchical asymmetric deep supervision.
• Model Variants: Two configurations were evaluated:
  • UltraSeg-130K: 0.13M parameters, targeting mobile real-time inference.
  • UltraSeg-500K: 0.5M parameters, targeting high-precision workstation deployment.

3. Experimental Setup and Results

The framework was validated across ten public datasets spanning six anatomical sites: breast (BUSI, UDIAT), thyroid (TN3K, DDTI), kidney (USKidney, URI-CADS), carotid artery (CCA), fetal assessment (HC18, PSFHS), and small-animal tumors (Mouse-GL261).

• Performance vs. Heavyweight Models:
  • Accuracy: UltraSeg-500K (0.5M params) achieved Dice scores comparable to or exceeding the 31M-parameter UNet and approached the 105M-parameter TransUNet across most datasets.
  • Boundary Precision: UltraSeg demonstrated superior boundary delineation (lower HD95) compared to both lightweight and heavyweight baselines, particularly in complex anatomical structures.
  • Zero-Shot Generalization: On external validation sets (UDIAT and DDTI), UltraSeg exhibited robust cross-dataset generalization, outperforming UNet and Att-UNet and matching or surpassing TransUNet in specific transfer scenarios.
• Inference Speed (GPU-Free Deployment):
  • Single-Core CPU: UltraSeg-130K achieved 51.7 FPS, and UltraSeg-500K achieved 27.8 FPS on an Intel i5 single-core CPU.
  • Mobile Device: On a refurbished Samsung Galaxy A53, UltraSeg-130K ran at 34.8 FPS and UltraSeg-500K at 16.1 FPS, enabling real-time guidance.
  • Comparison: In contrast, TransUNet and other heavyweight models failed to achieve real-time performance (<1 FPS) on these edge devices.
• Ablation Studies: Removing the EDB module caused significant performance drops (1.3–3.1% Dice). The ultrasound-specific dilation settings ([2, 4, 6]) consistently outperformed the original colonoscopy settings, validating the necessity of domain-specific receptive field design.

4. Key Contributions

1. GPU-Free Clinical AI: Demonstrated that clinical-grade ultrasound segmentation can be achieved on commodity CPUs and mobile devices without GPU dependency, aligning AI costs with the low cost of ultrasound hardware.
2. Ultra-Lightweight Architecture: Validated that sub-million parameter models (0.13M–0.5M) can match or exceed the performance of models with 30M–105M parameters in terms of Dice score and boundary accuracy.
3. Systematic Cross-Domain Validation: Provided extensive evidence of the framework's adaptability across six distinct anatomical sites and its robustness in zero-shot cross-dataset generalization.
4. Real-Time Edge Deployment: Established a technically feasible pathway for real-time, offline AI-assisted diagnosis in resource-constrained environments, including primary care clinics and mobile health units.

5. Significance and Claims

The paper argues that the current reliance on heavy, GPU-dependent AI models creates a barrier to health equity, excluding low-resource populations from the benefits of AI. By decoupling high-performance segmentation from specialized infrastructure, UltraSeg offers a paradigm shift:

• Democratization of Diagnostics: It enables primary care institutions and rural clinics to autonomously operate high-performance segmentation algorithms on local hardware, eliminating the need for expensive AI servers or cloud connectivity.
• Health Equity: The technology bridges the gap between device accessibility and diagnostic capability, ensuring that patients in resource-limited regions can access AI-assisted diagnostics comparable to developed areas.
• Sustainable Deployment: The framework addresses not just funding gaps but also infrastructure deficits (lack of IT staff, unreliable internet, power constraints) by offering an offline, energy-efficient solution compatible with portable ultrasound systems.

The authors conclude that while the technology is a specialized solution for standardized screening tasks and not a universal fix for all clinical complexities, it represents a critical step toward sustainable, locally owned AI in global healthcare.