An evaluation of Self-Supervised Pretraining for 3D Medical Image Segmentation
No Thumbnail Available
URL
Journal Title
Journal ISSN
Volume Title
Sähkötekniikan korkeakoulu |
Master's thesis
Authors
Date
2022-01-24
Department
Major/Subject
Autonomous Systems
Mcode
ELEC3055
Degree programme
Master's Programme in ICT Innovation
Language
en
Pages
6+77
Series
Abstract
In recent years, transfer learning has played an important role in numerous advancements in the field of Computer Vision in countless domains, including medical image analysis. This has lead transfer learning from natural image datasets, such as ImageNet, using standard large models and corresponding pretrained weights to become a common practice. However, some of the most prominent medical imaging modalities are 3D, including CT and MRI, which have to be reformulated in 2D in order to enable this transfer learning, inevitably resulting in the loss of 3D anatomical context and a compromise in performance. Self-supervision provides a different form of pretraining. In self-supervision, a label is generated from the data itself and used to train the model on a pretext task. This enables the model to train on unlabeled datasets, which are typically much larger than labeled datasets, that are still very similar in nature. Using self- supervision, the model can pretrain on domain specific 3D images, circumventing limitations of transfer learning. This thesis evaluates the impact of self-supervised pretraining on CT image segmentation. Using Models Genesis for self-supervised pretraining and nnU-Net for fine-tuning, it was shown that self-supervision can make the model more sample efficient showing a slight increase of 1.8% and 1.6% in mean dice score for models pretrained using an amount of unlabeled data equal to 10 times the amount of labeled data, and 4 times the amount of labeled data, respectively. Self-supervision is also shown to improve training efficiency, specifically for large unlabeled to labeled data ratios, reducing training time to reach 90% performance by 18% and 17% for a 10:1 and 4:1 ratio, respectively. Moreover, a new configuration for the out-painting augmentation of Models Genesis is shown to outperform the original.Description
Supervisor
Kannala, JuhoThesis advisor
Akram, SaadKeywords
medical image analysis, segmentation, self-supervised, deep learning