Accurate Medical Image Segmentation with Limited Annotations
OPEN ACCESS
Loading...
Author / Producer
Date
2022
Publication Type
Doctoral Thesis
ETH Bibliography
yes
Citations
Altmetric
OPEN ACCESS
Data
Rights / License
Abstract
Accurate image segmentation is important for many downstream clinical applications like diagnosis, surgery planning. In recent years, deep neural networks have been quite successful in providing high segmentation performance with fully supervised learning approaches. The bottleneck for such approaches is the requirement of a large amount of annotated training data. Acquiring such large labeled datasets is difficult for medical images as we need clinical experts for the annotations, which is expensive and time-consuming. Hence, this is not a preferred solution in clinical settings.
In this thesis, we focused on developing approaches to achieve high segmentation performance with limited annotations alongside leveraging the unlabeled data in training. The aim is to close the gap with the fully supervised approaches trained with large labeled datasets.
To achieve this, we propose the following three approaches:
In the first work, we propose a task-driven data augmentation method for learning with limited labeled data where the synthetic data generator is optimized for the segmentation task. The generator of the proposed method models intensity and shape variations using two sets of transformations, as additive intensity transformations and deformation fields. Both transformations are optimized using labeled as well as unlabeled examples in a semi-supervised framework.
In the second work, we propose an approach to pre-train a neural network with unlabeled data using a self-supervised learning approach to obtain a good network initialization that can be later fine-tuned to a target task. Here, we propose strategies for extending the contrastive learning framework, a self-supervised learning method for segmentation of volumetric medical images in the semi-supervised setting with limited annotations, by leveraging domain-specific and problem-specific cues. Specifically, we propose (1) novel contrasting strategies that leverage structural similarity across volumetric medical images (domain-specific cue) and (2) a local version of the contrastive loss to learn distinctive representations of local regions that are useful for per-pixel segmentation (problem-specific cue).
In the final work, we propose an end-to-end joint training framework to get high segmentation performance where we devise a contrastive loss to learn good local level features by exploiting semantic label information obtained from pseudo-labels of unlabeled images alongside limited annotated images. The pseudo-labels are rough segmentation mask estimates computed for the unlabeled images from a network initially trained with limited labeled images. In particular, we define the proposed loss to encourage similar representations for the pixels that have the same label and pseudo-label while being dissimilar to the representations of pixels from different classes in the dataset. We perform pseudo-label-based self-training and train the network by jointly optimizing the proposed contrastive loss on both labeled and unlabeled sets and segmentation loss on only the limited labeled set.
All the above-proposed approaches have been evaluated extensively on three MRI/CT datasets. They yielded better results than the compared data augmentation, semi-supervised, and self-supervised learning approaches.
Permanent link
Publication status
published
External links
Editor
Contributors
Examiner : Konukoglu, Ender
Examiner : Rueckert, Daniel
Examiner : Mansi, Tommaso
Book title
Journal / series
Volume
Pages / Article No.
Publisher
ETH Zurich
Event
Edition / version
Methods
Software
Geographic location
Date collected
Date created
Subject
Deep learning; Data augmentation; Semi supervised learning; Self-supervised learning
Organisational unit
09579 - Konukoglu, Ender / Konukoglu, Ender