Oude Vrielink, Jeroen (2024) Understanding Self-Supervised Pretraining Methods for Label-Scarce Ankle Fracture Classification. Master's Thesis / Essay, Artificial Intelligence.
|
Text
mAI2024OudeVrielinkJ.pdf Download (8MB) | Preview |
|
![[img]](https://fse.studenttheses.ub.rug.nl/style/images/fileicons/text.png) |
Text
toestemming.pdf Restricted to Registered users only Download (506kB) |
Abstract
Acquiring and annotating medical image data is a time-consuming process that requires specialized expertise and can be hindered by privacy and ethical concerns, causing a bottleneck for supervised deep learning methods in the medical imaging field. Self-supervised learning (SSL) offers a promising solution to this challenge by learning informative representations from unlabeled data. We investigate the viability of SSL pretraining for downstream ankle fracture classification, aiming to address the challenge of label scarcity in medical imaging. We pretrain a ResNet-50 backbone with one masked image modeling method (SparK) and five contrastive learning methods (MoCo v2/v3, VICReg(L), DINO) on our curated novel Large unlabeled X-ray (LuXry) dataset. Our findings demonstrate that DINO pretraining achieves an 8.8% improvement over the supervised baseline, while the other SSL methods lead to decreased performance. Through ablation studies, singular value decomposition analysis, and visualization of latent representations, we suggest a transferability issue of SSL features to the downstream task and highlight fundamental properties contributing to DINO’s success. Additionally, we propose a novel projection head architecture for DINO and a content-aware patch masking strategy for SparK, demonstrating their viability for feature learning
| Item Type: | Thesis (Master's Thesis / Essay) |
|---|---|
| Supervisor name: | Mohades Kasaei, S.H. |
| Degree programme: | Artificial Intelligence |
| Thesis type: | Master's Thesis / Essay |
| Language: | English |
| Date Deposited: | 25 Jul 2024 12:13 |
| Last Modified: | 30 Jul 2024 14:18 |
| URI: | https://fse.studenttheses.ub.rug.nl/id/eprint/33515 |
Actions (login required)
 |
View Item |