TY - JOUR
T1 - Automatic coarse-to-refinement-based ultrasound prostate segmentation using optimal polyline segment tracking method and deep learning
AU - Peng, Tao
AU - Xu, Daqiang
AU - Tang, Caiyin
AU - Zhao, Jing
AU - Shen, Yuntian
AU - Yang, Cong
AU - Cai, Jing
N1 - Publisher Copyright:
© 2023, The Author(s), under exclusive licence to Springer Science+Business Media, LLC, part of Springer Nature.
PY - 2023
Y1 - 2023
N2 - Automatic segmentation of the prostate in transrectal ultrasound (TRUS) images provides useful information for prostate cancer diagnosis and treatment. However, boundaries between the prostate and other tissues are often absent or ill defined in TRUS images, which means that automatic segmentation of the prostate in TRUS images is highly challenging. In this study, we attempted to overcome these challenges by developing a novel method we termed “automatic prostate segmentation” (Auto-ProSeg) that is capable of effectively segmenting the prostate in TRUS images. Auto-ProSeg comprises two steps: the first step is a preprocessing step that uses attention U-Net to extract approximate prostate contours automatically; then, in the second step, the approximate prostate contours are optimized via a modified principal curve-based method linked to an evolutionary neural network, whereby a mathematical mapping formula based on the parameters of an enhanced evolutionary neural network is used to generate smooth prostate contours. Our results illustrate that Auto-ProSeg exhibits better prostate segmentation performance than other recently developed methods: the average Dice similarity coefficient and Jaccard similarity coefficient (Ω) of Auto-ProSeg-generated prostate contours against ground truths were 94.2% ± 3.2% and 93% ± 3.7%, whereas those for other state-of-the-art fully automatic segmentation methods were approximately 90% ± 5% and 89% ± 6%, respectively.
AB - Automatic segmentation of the prostate in transrectal ultrasound (TRUS) images provides useful information for prostate cancer diagnosis and treatment. However, boundaries between the prostate and other tissues are often absent or ill defined in TRUS images, which means that automatic segmentation of the prostate in TRUS images is highly challenging. In this study, we attempted to overcome these challenges by developing a novel method we termed “automatic prostate segmentation” (Auto-ProSeg) that is capable of effectively segmenting the prostate in TRUS images. Auto-ProSeg comprises two steps: the first step is a preprocessing step that uses attention U-Net to extract approximate prostate contours automatically; then, in the second step, the approximate prostate contours are optimized via a modified principal curve-based method linked to an evolutionary neural network, whereby a mathematical mapping formula based on the parameters of an enhanced evolutionary neural network is used to generate smooth prostate contours. Our results illustrate that Auto-ProSeg exhibits better prostate segmentation performance than other recently developed methods: the average Dice similarity coefficient and Jaccard similarity coefficient (Ω) of Auto-ProSeg-generated prostate contours against ground truths were 94.2% ± 3.2% and 93% ± 3.7%, whereas those for other state-of-the-art fully automatic segmentation methods were approximately 90% ± 5% and 89% ± 6%, respectively.
KW - Attention gate Unet
KW - Evolution neural network
KW - Mathematical mapping formula
KW - Optimal polyline segment tracking method
KW - Ultrasound prostate segmentation
UR - http://www.scopus.com/inward/record.url?scp=85160736470&partnerID=8YFLogxK
U2 - 10.1007/s10489-023-04676-4
DO - 10.1007/s10489-023-04676-4
M3 - Journal article
AN - SCOPUS:85160736470
SN - 0924-669X
JO - Applied Intelligence
JF - Applied Intelligence
ER -