We gratefully acknowledge support from
the Simons Foundation and member institutions.
Full-text links:

Download:

Current browse context:

eess.IV

Change to browse by:

References & Citations

Bookmark

(what is this?)
CiteULike logo BibSonomy logo Mendeley logo del.icio.us logo Digg logo Reddit logo ScienceWISE logo

Electrical Engineering and Systems Science > Image and Video Processing

Title: Unsupervised Anomaly Segmentation using Image-Semantic Cycle Translation

Abstract: The goal of unsupervised anomaly segmentation (UAS) is to detect the pixel-level anomalies unseen during training. It is a promising field in the medical imaging community, e.g, we can use the model trained with only healthy data to segment the lesions of rare diseases. Existing methods are mainly based on Information Bottleneck, whose underlying principle is modeling the distribution of normal anatomy via learning to compress and recover the healthy data with a low-dimensional manifold, and then detecting lesions as the outlier from this learned distribution. However, this dimensionality reduction inevitably damages the localization information, which is especially essential for pixel-level anomaly detection. In this paper, to alleviate this issue, we introduce the semantic space of healthy anatomy in the process of modeling healthy-data distribution. More precisely, we view the couple of segmentation and synthesis as a special Autoencoder, and propose a novel cycle translation framework with a journey of 'image->semantic->image'. Experimental results on the BraTS and ISLES databases show that the proposed approach achieves significantly superior performance compared to several prior methods and segments the anomalies more accurately.
Subjects: Image and Video Processing (eess.IV); Computer Vision and Pattern Recognition (cs.CV)
Cite as: arXiv:2103.09094 [eess.IV]
  (or arXiv:2103.09094v1 [eess.IV] for this version)

Submission history

From: Chenxin Li [view email]
[v1] Tue, 16 Mar 2021 14:15:30 GMT (1380kb,D)

Link back to: arXiv, form interface, contact.