Introduction
Standard ultrasound probes have a limited field of view (FOV), which can be restrictive when clinicians need to visualize large organs or long lesions. Extended Field of View Ultrasound (EFOV-US) addresses this by stitching multiple frames into a single panoramic image. However, ultrasound images are often characterized by speckle noise, low contrast, and low resolution, making it difficult for computer vision algorithms to identify the "feature points" necessary for accurate alignment. Sparse or inaccurate feature points lead to stitching artifacts, blurring, and geometric distortions.
Proposed Methodology
To overcome these challenges, researchers developed a Dual-enhanced EFOV-US pipeline. This approach enhances the image in two distinct stages before the final registration process.
1. Gray-scale Enhancement
The first stage targets speckle noise and low contrast. The process involves inverting the image, fitting a quadratic curve to the grayscale histogram to identify a noise threshold, and then normalizing the histogram. This "cleans" the image and highlights structural details that were previously obscured.
2. GAN-based Super-Resolution
Following denoising, the image undergoes super-resolution using a Generative Adversarial Network (GAN), specifically a variation of SRResNet. Unlike standard interpolation, the GAN generates high-frequency details, increasing the resolution and creating a sharper image with more detectable landmarks.
3. Registration and Fusion
For the stitching phase, the system uses SIFT (Scale-Invariant Feature Transform) for point extraction and GLOH (Gradient Location and Orientation Histogram) for description. GLOH was selected for its robustness to the specific types of blur and noise found in medical imaging. Finally, RANSAC filters out outlier matches, and a weighted average method fuses the images into a seamless panorama.
Experimental Results
The method was evaluated on a clinical dataset of 84 ultrasound images of the Common Carotid Artery (CCA). The "Dual-Enhanced" approach significantly outperformed traditional methods:
- Feature Extraction: The number of successfully matched feature points increased from 108 in the original images to 496 in the dual-enhanced versions.
- Image Sharpness: The proposed method achieved an Average Gradient of 5.6258, more than double the 2.2001 of standard EFOV-US.
- Information Content: The Standard Deviation (measuring detail and contrast) reached 92, compared to 37 in the baseline model.
- Efficiency: The total calculation time was reduced to approximately 3 seconds, bringing the technology closer to real-time clinical application.
Conclusion
The Dual-Enhanced registration method demonstrates that significant "hidden" information can be recovered from noisy ultrasound frames through effective preprocessing. By combining denoising with GAN-based super-resolution, the system generates a much larger pool of reliable feature points, leading to more stable and accurate panoramic images. Future work will focus on optimizing the GAN for diverse clinical environments and achieving true real-time stitching.
Summary updated