Intelligent Ultrasound Image Segmentation Using Adaptive Coupled Fractional Deformable Model

Precise and automated ultrasound image segmentation is essential for improving computer-aided disease diagnosis. However, image degradations such as low contrast, intensity inhomogeneity, and speckle noise often obscure tissue details and reduce the accuracy of segmentation methods. While deformable models are theoretically explanatory and do not require large datasets, they are sensitive to initial contour placement. This article introduces a coarse-to-fine segmentation approach integrating deep learning with an adaptive coupled fractional deformable model (CFDM). Initially, the YOLOv9 (you only look once version 9) model is trained to generate coarse segmentations due to its boundary delineation efficiency. However, YOLOv9 struggles with degraded images and blurred edges, leading to coarse boundaries. To address this, the CFDM refines these boundaries by using adaptive fractional order, correcting inhomogeneity, and despeckling to improve segmentation accuracy. The well-posedness of the hybrid CFDM (HCFDM) is demonstrated, and experiments on breast ultrasound images of benign and malignant tumors show that the HCFDM outperforms five integer order deformable models, five fractional active contour models, four deep learning models, and two hybrid models. This makes it highly effective for automatic ultrasound image segmentation in clinical applications.