Tympanic membrane segmentation in otoscopic images based on fully convolutional network with active contour loss

Van Truong Pham, Thi Thao Tran, Pa Chun Wang, Men Tzung Lo

Research output: Contribution to journalArticlepeer-review

3 Scopus citations

Abstract

This paper presents a method to automatically segment tympanic membranes (TMs) from video-otoscopic images based on the deep learning approach. The paper introduces a hybrid loss function combining the Dice loss and active contour loss to the fully convolutional network. By this way, the proposed model takes into account the Dice similarity and the desired boundary contour information including the contour length as well as regions inside and outside the contour during learning. The proposed loss function is then applied to the fully convolutional network for tympanic membrane segmentation. We evaluate the proposed approach on TMs data set which includes 1139 otoscopic images from patients diagnosed with and without otitis media. Experimental results show that the proposed deep learning model achieves an average Dice similarity coefficient of 0.895, a mean Hausdorff distance of 19.189, and average perpendicular distance of 6.429, that outperforms other state-of-the-art methods.

Original languageEnglish
Pages (from-to)519-527
Number of pages9
JournalSignal, Image and Video Processing
Volume15
Issue number3
DOIs
StatePublished - Apr 2021

Keywords

  • Active contour model
  • Deep learning
  • Image segmentation
  • Otitis media
  • Tympanic membranes segmentation

Fingerprint

Dive into the research topics of 'Tympanic membrane segmentation in otoscopic images based on fully convolutional network with active contour loss'. Together they form a unique fingerprint.

Cite this