Prostate Segmentation via Dynamic Fusion Model


Ocal H., BARIŞÇI N.

ARABIAN JOURNAL FOR SCIENCE AND ENGINEERING, vol.47, no.8, pp.10211-10224, 2022 (Peer-Reviewed Journal) identifier identifier

  • Publication Type: Article / Article
  • Volume: 47 Issue: 8
  • Publication Date: 2022
  • Doi Number: 10.1007/s13369-021-06502-w
  • Journal Name: ARABIAN JOURNAL FOR SCIENCE AND ENGINEERING
  • Journal Indexes: Science Citation Index Expanded, Scopus, Aerospace Database, Communication Abstracts, Metadex, Pollution Abstracts, zbMATH, Civil Engineering Abstracts
  • Page Numbers: pp.10211-10224
  • Keywords: Deep learning, Prostate segmentation, Unet, ResNet, Dynamic case-wise focal Tversky Loss, MRI

Abstract

Nowadays, many different methods are used in diagnosing prostate cancer. Among these methods, MRI-based imaging methods provide more precise information than other methods by obtaining the prostate's image from different angles (axial, sagittal, coronal). However, manually segmenting these images is very time-consuming and laborious. Besides, another challenge is the inhomogeneous and inconsistent appearance around the prostate borders, which is essential for cancer diagnosis. Nowadays, scientists are working intensively on deep learning-based techniques to identify prostate boundaries more efficiently and with high accuracy. In this study, a dynamic fusion architecture is proposed. For the fusion model, the Unet + Resnet3D and Unet + Resnet2D models were fused. 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. Comparative analyzes show that the advantages and robustness of our method are superior to state-of-the-art approaches.