Delineation of Prostate Boundary from Medical Images via a Mathematical Formula-Based Hybrid Algorithm

The precise extraction of the contour of prostate on transrectal ultrasound (TRUS) is crucial for the diagnosis and treatment of prostate tumor. Due to the relatively low signal-to-noise ratio (SNR) of TRUS images and the potential of imaging artifacts, accurate contouring of the prostate from TRUS images has been a challenging task. This paper proposes four strategies to achieve higher precision of segmentation on TRUS images. Firstly, a modified principal curve-based algorithm is used to obtain the data sequence, with a small amount of prior point information adopted for coarse initialization. Secondly, an evolution neural network is devised to find an optimal network. Thirdly, a fractional-order-based network is trained with the data sequence as input, resulting in a decreased model error and increased precision. Finally, the parameters of a fractional-order-based neural networkwere utilized to construct an interpretable and smoothmathematical equation of the organ border. The Dice similarity coefficient (DSC), Jaccard similarity coefficient (OMG), and accuracy (ACC) of model outputs against ground-truths were 95.9 +/- 2.3%, 94.9 +/- 2.4%, and 95.3 +/- 2.2%, respectively. The results of our method outperform several popular state-of-the-art segmentation methods.