A novel prostate segmentation method: triple fusion model with hybrid loss


Ocal H., BARIŞÇI N.

NEURAL COMPUTING & APPLICATIONS, vol.34, no.16, pp.13559-13574, 2022 (SCI-Expanded) identifier identifier

  • Publication Type: Article / Article
  • Volume: 34 Issue: 16
  • Publication Date: 2022
  • Doi Number: 10.1007/s00521-022-07188-3
  • Journal Name: NEURAL COMPUTING & APPLICATIONS
  • Journal Indexes: Science Citation Index Expanded (SCI-EXPANDED), Scopus, Academic Search Premier, PASCAL, Applied Science & Technology Source, Biotechnology Research Abstracts, Compendex, Computer & Applied Sciences, Index Islamicus, INSPEC, zbMATH
  • Page Numbers: pp.13559-13574
  • Keywords: Deep learning, Prostate segmentation, Triple fusion model, Weighted focal Tversky Loss, MAGNETIC-RESONANCE, ULTRASOUND
  • Gazi University Affiliated: Yes

Abstract

Early and rapid diagnosis of prostate cancer, the horsehead disease among men, has become increasingly important. Nowadays, many methods are used in the early diagnosis of prostate cancer. Compared to other imaging methods, magnetic resonance imaging (MRI) based on prostate gland imaging is preferred because angular imaging (axial, sagittal, and coronal) provides precise information. But diagnosing the disease from these MR images is time-consuming. For example, imaging differences between MR devices for prostate segmentation and inhomogeneous and inconsistent prostate appearance are significant challenges. Because of these segmentation difficulties, manual segmentation of prostate images is challenging. In recent years, computer-aided intelligent architectures (deep learning-based architecture) have been used to overcome the manual segmentation of prostate images. These architectures can now perform manual prostate segmentation in seconds that used to take days thanks to their end-to-end automatic deep convolutional neural networks (DCNN). Inspired by the studies mentioned above, this study proposes a novel DCNN approach for prostate segmentation by combining ResUnet 2D with residual blocks and Edge Attention Vnet 3D architectures. In addition, the weighted foal Twersky loss function, which was proposed for the first time, significantly increased the architecture's performance. Evaluation experiments were performed on the MICCAI 2012 Prostate Segmentation Challenge Dataset (PROMISE12) and the NCI-ISBI 2013(NCI_ISBI-13) Prostate Segmentation Challenge Dataset. As a result of the tests performed, Dice scores of 91.92 and 91.15% in the whole prostate volume were obtained in the Promise 12 and NCI_ISBI 13 datasets, respectively. Comparative analyses show that the advantages and robustness of our method are superior to the state-of-the-art approaches.