Improved Kidney Outlining in Ultrasound Images by Combining Deep Learning Semantic Segmentation with Conventional Active Contour

Computer-aided diagnosis systems are used to process kidney ultrasound images with the goal of providing the clinicians with objective, computer-based analyses to improve the diagnosis accuracy. A crucial requirement to develop these systems is to accurately outline the kidney in ultrasound images. The majority of the computer-based methods that are proposed for outlining the kidney in ultrasound images are based on conventional segmentation approaches, such as active contour algorithms. Recently, deep learning semantic segmentation models have been employed to outline the kidney in ultrasound images. However, limited work has been made to investigate the possibility of combining conventional segmentation methods with the recently introduced deep learning semantic segmentation models to achieve effective outlining of the kidney. This study investigates the feasibility of combining deep learning semantic segmentation with a conventional active contour segmentation algorithm with the goal of improving the outlining of the kidney in ultrasound images. In particular, the performance of three deep learning semantic segmentation models, which are configured to use different backbone convolutional neural networks, has been studied in terms of their ability to detect the pixels that belong to the kidney in the ultrasound image. The best performing semantic segmentation approach is used to generate an initial outline for the kidney. Furthermore, a conventional edge-based active contour segmentation algorithm is used to refine and improve the initial kidney outline. The results indicate that our proposed approach for combining the deep learning semantic segmentation with the conventional edge-based active contour segmentation has achieved effective kidney outlining with Precision, Recall, F1-score, Accuracy, and Jaccard values of 94.2%, 89.6%, 91.0%, 98.4%, and 84.5%, respectively.