Fully automatic segmentation on prostate MR images based on cascaded fully convolution network

• Published in: Journal of magnetic resonance imaging Vol. 49; no. 4; pp. 1149 - 1156
• Main Authors: Zhu, Yi, Wei, Rong, Gao, Ge, Ding, Lian, Zhang, Xiaodong, Wang, Xiaoying, Zhang, Jue
• Format: Journal Article
• Language: English
• Published: United States Wiley Subscription Services, Inc 01.04.2019
• Subjects: Algorithms; Cancer; cascaded fully convolutional network; Contours; Convolution; Field strength; fully automatic segmentation; Image processing; Image segmentation; Magnetic resonance imaging; Medical imaging; Population (statistical); Prostate cancer; prostatic peripheral zone; Shape; Statistical analysis; Statistical tests; the ROI of prostate; prostatic peripheral zone; fully automatic segmentation; the ROI of prostate; cascaded fully convolutional network
• DOI: 10.1002/jmri.26337
• ISSN: 1053-1807; 1522-2586

Background Computer‐aided diagnosis (CAD) can aid radiologists in quantifying prostate cancer, and MRI segmentation plays an essential role in CAD applications. Clinical experience shows that prostate cancer occurs predominantly in the peripheral zone (PZ) and there exist different evaluation criteria for different regions in the Prostate Imaging Reporting and Data System (PI‐RADS). Purpose To develop a fully automatic approach to segmenting the prostate outer contour and the PZ contour with high efficacy. Population In all, 163 subjects, including 61 healthy subjects and 102 prostate cancer patients. For each subject, all slices that contained prostate gland in diffusion‐weighted images (DWIs) and T2‐weighted images (T2WIs) were selected as the datasets. Field Strength T2‐weighted, DWI at 3.0T. Assessment The computer‐generated segmentation results were compared with the manual outlining results that were depicted by two experts with more than 5 years' experience. Dice similarity coefficient (DSC), false‐positive rate (FPR), and false‐negative rate (FNR) were used to compared the algorithm and manual segmentation results. Statistical Tests A paired t‐test was adopted for comparison between our method and classical U‐Net segmentation methods. Results The mean DSC was 92.7 ± 4.2% for the total whole prostate gland and 79.3 ± 10.4% for the total peripheral zone. Compared with classical U‐Net segmentation methods, our segmentation precision was significantly higher (P < 0.001). Data Conclusion By cropping the region of interest and cascading the two networks, our method balances the positive and negative sample gradually, and results in higher segmentation accuracy. This fully automatic strategy could provide satisfactory performance in prostate DWIs and T2WIs‐based image segmentation. Level of Evidence: 2 Technical Efficacy Stage 1 J. Magn. Reson. Imaging 2019;49:1149–1156.