Skip to main content

LUP Student Papers

LUND UNIVERSITY LIBRARIES

Semi-supervised Lesion Detection via Image Restoration

Kronander, Jonatan LU (2020) In Master’s Theses in Mathematical Sciences FMAM05 20201
Mathematics (Faculty of Engineering)
Abstract
Lesion detection is a critical task for medical image understanding. While the problem has been widely addressed in a supervised semantic segmentation manner, the problem clinically appears more similar to novelty detection with few or no annotations for the lesion. The reason is two-fold: 1) it is intuitively easier to collect large dataset from healthy individuals than that from a specific type of lesion individuals, 2) clinicians are generally interested in any abnormalities regardless of its type. This makes unsupervised methods more attractive solutions. Works such as AnoGAN and VAE with image restoration offer practical ways to localise lesions by training only on healthy data. However, for the same type of lesion, an obvious... (More)
Lesion detection is a critical task for medical image understanding. While the problem has been widely addressed in a supervised semantic segmentation manner, the problem clinically appears more similar to novelty detection with few or no annotations for the lesion. The reason is two-fold: 1) it is intuitively easier to collect large dataset from healthy individuals than that from a specific type of lesion individuals, 2) clinicians are generally interested in any abnormalities regardless of its type. This makes unsupervised methods more attractive solutions. Works such as AnoGAN and VAE with image restoration offer practical ways to localise lesions by training only on healthy data. However, for the same type of lesion, an obvious performance gap exists between unsupervised and supervised methods. In this work, we intend to provide supervision with a small number of lesion data to the unsupervised method with the aim to narrow the gap. The method is an extension of the unsupervised method of VAE with MAP-based image restoration. In more details, we train an U-Net on the few examples to predict the likelihood term and impose the supervision with annotated lesions such that the restoration only occurs for the lesion pixels. We train the unsupervised method on T2-weighted images of healthy individuals of Cam-CAN dataset and provide a small annotated dataset consisting of a few subjects from BraTS dataset, and test on an unseen subset of BraTS. With the addition of the few examples, the method shows an improvement over the unsupervised method while the gap with the supervised is narrowed but still exists. (Less)
Please use this url to cite or link to this publication:
author
Kronander, Jonatan LU
supervisor
organization
course
FMAM05 20201
year
type
H2 - Master's Degree (Two Years)
subject
publication/series
Master’s Theses in Mathematical Sciences
report number
LUTFMA-3427-2020
ISSN
1404-6342
other publication id
2020:E69
language
English
id
9027679
date added to LUP
2020-09-18 10:49:02
date last changed
2020-09-18 10:49:02
@misc{9027679,
  abstract     = {{Lesion detection is a critical task for medical image understanding. While the problem has been widely addressed in a supervised semantic segmentation manner, the problem clinically appears more similar to novelty detection with few or no annotations for the lesion. The reason is two-fold: 1) it is intuitively easier to collect large dataset from healthy individuals than that from a specific type of lesion individuals, 2) clinicians are generally interested in any abnormalities regardless of its type. This makes unsupervised methods more attractive solutions. Works such as AnoGAN and VAE with image restoration offer practical ways to localise lesions by training only on healthy data. However, for the same type of lesion, an obvious performance gap exists between unsupervised and supervised methods. In this work, we intend to provide supervision with a small number of lesion data to the unsupervised method with the aim to narrow the gap. The method is an extension of the unsupervised method of VAE with MAP-based image restoration. In more details, we train an U-Net on the few examples to predict the likelihood term and impose the supervision with annotated lesions such that the restoration only occurs for the lesion pixels. We train the unsupervised method on T2-weighted images of healthy individuals of Cam-CAN dataset and provide a small annotated dataset consisting of a few subjects from BraTS dataset, and test on an unseen subset of BraTS. With the addition of the few examples, the method shows an improvement over the unsupervised method while the gap with the supervised is narrowed but still exists.}},
  author       = {{Kronander, Jonatan}},
  issn         = {{1404-6342}},
  language     = {{eng}},
  note         = {{Student Paper}},
  series       = {{Master’s Theses in Mathematical Sciences}},
  title        = {{Semi-supervised Lesion Detection via Image Restoration}},
  year         = {{2020}},
}