Purpose: Transrectal ultrasound (TRUS) is routinely used in ultrasound-guided brachytherapy of prostate cancer for treatment planning and verification. Manual delineation of the prostate volume is tedious, time-consuming, and subjective. The purpose of this work is to develop a prostate segmentation hybrid method using an improved redefined closed k-Segment principal curve and improved differential evolution-based artificial neural network.
Methods: TRUS scans of 55 brachytherapy patients obtained from the Vinno 70 Lab system were analyzed. We developed a novel method for TRUS prostate segmentation, using only a small number (< 15%) of manually delineated points as the approximate initialization. This method contains three subnetworks: the first one uses an improved principal curve-based method to obtain the data sequences consisting of data points and projection indexes; the second one is using improved Memory-based and Cuckoo Search-based Differential Evolution (MCSDE) method for preliminarily searching for the optimal model of the machine learning method; lastly, we proposed a smooth mathematical model of prostate contour, which was denoted by the parameters of machine learning method.
Results: Preliminary results show that the average Dice Similarity Coefficient (DSC), Jaccard Similarity Coefficient (Ω), and Accuracy (ACC) is 95.8%, 94.3%, and 95.4% on the testing dataset (15 patients). Meanwhile, the DSC, Ω, and ACC of the proposed method are as high as 93.3%, 91.9%, and 93%, respectively, at the influence of Gaussian noise (standard deviation of Gaussian function σ=50). Although σ increases from 10 to 50, the DSC, Ω, and ACC fluctuate about 2.5% at most, demonstrating the good robustness of our method. Compared with the competing models, our method has the best performance.
Conclusion: Both qualitative and quantitative experimental results demonstrate that our hybrid method of prostate segmentation achieved superior performance compared with several state-of-the-art methods. Meanwhile, our method has good robustness to deal with the noise data effectively.
Funding Support, Disclosures, and Conflict of Interest: This work is partly supported by GRF 151022/19M.