Ultrasound lesion segmentation using clinical knowledge-driven constrained level set.

Ultrasound lesion segmentation remains a difficult challenge due to image characteristics like shadowing, speckle, and heterogeneity. Region-based level set methods offer benefits but struggle with these issues. Physicians often mark lesion axes for measurements, creating four image markers. This clinical practice has not been fully integrated into segmentation algorithms. Prior research has shown that manual segmentation is time-consuming and inconsistent. No prior work had resolved how to use these markers as constraints in automated methods. This gap motivated the development of a marker-driven segmentation approach. The goal is to improve accuracy and efficiency by leveraging clinical data. The study addresses the need for real-time, reliable segmentation tools.

The study aimed to improve ultrasound lesion segmentation by incorporating clinical markers into the level set method. Physicians typically mark lesion axes, creating four markers in an image. These markers were used as prior knowledge in the segmentation process. The objective was to ensure the contour evolution stays aligned with the markers. The researchers proposed a constrained energy function to guide the contour. This approach was designed to handle ultrasound-specific challenges. The method was tested on a large dataset of breast lesion images. The goal was to achieve high accuracy and real-time performance.

The method begins with template-matching to detect lesion markers in ultrasound images. B-Spline fitting is then used to create an initial contour from the four markers. A constrained energy function is added to the region-based level set framework. This function ensures the contour remains close to the markers during evolution. The algorithm operates in a multi-resolution scheme for efficiency. This step reduces computational load while maintaining accuracy. The method was validated using 308 manually segmented breast lesion images. Performance metrics included the Dice similarity coefficient and processing speed.

The proposed method achieved a Dice similarity coefficient of 89.49 ± 4.76% on 308 images. This score was compared to manually segmented boundaries. The algorithm ran in real-time, making it suitable for clinical use. The constrained energy function effectively guided contour evolution. The multi-resolution approach improved computational efficiency. The method outperformed standard level set techniques in accuracy. Template-matching reliably detected the four markers. B-Spline fitting provided a stable initial contour for segmentation.

The study demonstrated that using clinical markers improves segmentation accuracy. The constrained energy function ensured contour alignment with markers. Real-time performance was achieved through a multi-resolution scheme. The method was validated on a large dataset of breast lesions. The results suggest this approach is suitable for clinical workflows. Physicians can benefit from faster and more accurate segmentation. The method handles ultrasound-specific challenges like speckle and shadowing. The authors propose this technique as a practical solution for lesion segmentation.