This thesis explores using self- and semi-supervised learning for multi-class segmentation of distal radius fractures with limited labeled data, highlighting modest gains and dataset challenges.
Python, PyTorch, Self-Supervised Learning, Semi-Supervised Learning
The goal of this thesis is to improve the segmentation of distal radius fractures in medical imaging by exploring advanced machine learning techniques. Specifically, the research aims to enhance the accuracy and efficiency of multi-class segmentation on 3D CT scans using semi- and self-supervised learning approaches, combined with vision transformers and foundation models. By leveraging these techniques, the study seeks to overcome the limitations of traditional supervised learning methods, particularly in scenarios where labeled data is scarce.
Segmenting wrist fractures on 3D CT scans according to Dr. Marco Keller's classification requires extensive expert knowledge and a significant amount of time. This complexity results in a limited dataset of only 60 labeled CT scans available for supervised training, making it challenging to achieve high performance with conventional methods. To address this, the study incorporates an additional 452 unlabeled CT scans for self-supervised pre-training.
The results of the thesis show modest improvements in the segmentation performance of distal radius fractures when using semi- and self-supervised learning methods compared to traditional supervised techniques. However, the small dataset size and class imbalance present significant challenges, limiting the effectiveness of these advanced methodologies. The findings underscore the importance of larger, more balanced datasets to fully make use of self-supervised learning.
Project Duration: March 2024 - August 2024, Effort: 720h, Team Size: 2 people
Dr. med. Marco Keller, Kantonsspital Baselland
Yves Bächtiger, Lukas Weber
Dr. Cédric Huwyler
E-Mail: cedric.huwyler@fhnw.ch
Prof. Dr. Martin Melchior
E-Mail: martin.melchior@fhnw.ch
