Hybrid K-means and CNN Refinement for Kidney Images Segmentation

Abstract

This study aims to present a two-stage framework combining artificial intelligence algorithms, including the efficient K-means algorithm and a lightweight convolutional neural network (CNN) algorithm, which provides high accuracy for precise segmentation of Ultrasound images and obtaining more accurate data. Image segmentation is a fundamental process for image analysis and obtaining more precise information; it is one of the best methods used in image processing. The K-means algorithm generates an initial mask (an initial image segmentation). The CNN algorithm, a neural network, then refines this mask to recover the precise boundaries of objects within the overall image, thus identifying the image's edges without loss of features and pinpointing the location for diagnosis. The results obtained from the Ultrasound images used on the BSD500 and PASCAL VOC 2012 datasets showed a clear improvement compared to using K-means alone, as the CNN algorithm proved its worth in obtaining more accurate and clear data, and achieving performance that rivals the latest deep learning-based methods, while maintaining real-time execution speed at the same time.