Paper Abstract and Keywords |
Presentation |
2022-01-26 10:13
[Short Paper]
Abnormality Detection for Covid-19 Chest CT Images by Dimensionality Reduction Based on Contrastive Learning Hiroki Tobise, Kugler Mauricio, Tatsuya Yokota (NITech), Masahiro Hashimoto (Keio Univ.), Yoshito Otake (NAIST), Toshiaki Akashi (Juntendo Univ.), Akinobu Shimizu (TUAT), Hidekata Hontani (NITech) MI2021-53 |
Abstract |
(in Japanese) |
(See Japanese page) |
(in English) |
In this article, we propose a method that detects anomaly regions in chest CT images for the aid of Covid-19 diagnosis. Employing an approach for constructing a 1-class classifier based on the probability distribution of patch images of normal cases, we can relax the unbalance of the training data between different classes. The probability distribution should be estimated not in the patch image space but in a low-dimensional space in which we can estimate the similarity between patch images by referring to the Euclid distance between them. We therefore employ a contrastive-loss-based self-supervised learning method for the dimensionality reduction. The contrastive-loss is useful for realizing the projection invariant to the operations defined by users. We obtain a projection of patch images that is invariant against translation and flipping. Some experimental results are reported in this presentation. |
Keyword |
(in Japanese) |
(See Japanese page) |
(in English) |
chest CT images / Covid-19 / anomaly detection / neural network / Contrastive Learning / Normalizing Flows / / |
Reference Info. |
IEICE Tech. Rep., vol. 121, no. 347, MI2021-53, pp. 41-42, Jan. 2022. |
Paper # |
MI2021-53 |
Date of Issue |
2022-01-18 (MI) |
ISSN |
Online edition: ISSN 2432-6380 |
Copyright and reproduction |
All rights are reserved and no part of this publication may be reproduced or transmitted in any form or by any means, electronic or mechanical, including photocopy, recording, or any information storage and retrieval system, without permission in writing from the publisher. Notwithstanding, instructors are permitted to photocopy isolated articles for noncommercial classroom use without fee. (License No.: 10GA0019/12GB0052/13GB0056/17GB0034/18GB0034) |
Download PDF |
MI2021-53 |
|