Medical image segmentation with scattering-based self-supervised learning

Although deep learning–based methods have achieved significant success in image segmentation problems in recent years, the supervised learning approaches underlying this success require large amounts of pixel-level annotated data. Particularly in the field of medical imaging, the need for expert knowledge in the annotation process, along with its time-consuming and costly nature, makes the curation of large-scale labeled datasets considerably challenging. Therefore, the development of learning approaches that are less dependent on labeled data has become a critical requirement in medical image segmentation. In this thesis, a scattering-based self-supervised learning approach is proposed to reduce the dependency on labeled data in medical image segmentation problems. In the proposed method, fixed-filter Wavelet Scattering Networks (WSN), which provide mathematically grounded properties such as translation invariance and stability to small deformations, and Parametric Scattering Networks (PSN) with learnable wavelet parameters are integrated into Bootstrap Your Own Latent (BYOL), a non-contrastive self-supervised learning framework. The encoders obtained after self-supervised pretraining are evaluated through supervised fine-tuning within a U-Net architecture. The proposed approach was evaluated through experiments conducted under different labeled data scenarios on the ACDC dataset, which consists of cardiac magnetic resonance images, and the CHD dataset, composed of computed tomography images of congenital heart disease. The obtained results demonstrate that scattering-based self-supervised initialization strategies provide higher label efficiency under limited labeled data conditions compared to random initialization, ImageNet pretraining, and the standard BYOL approach. In particular, the PSA-based approach achieves the highest segmentation performance for anatomically challenging structures. These findings indicate that the joint use of scattering transforms and self-supervised learning offers an effective solution for label-efficient medical image segmentation.