In this thesis we deal with the problem of segmenting single cell instances in microscopic images.
For specific datasets, manual labelling of instances is an expensive and time-consuming process, so our aim is to improve the segmentation results of the final model using self-supervised learning.
We used the Cellpose approach for the segmentation task and the masked autoencoder approach for the pre-training task.
The model architecture is based on the U-Net and was tested with a several convolutional and a transformer-based encoder.
We compared the results of pre-training using the masked autoencoder approach on domain-specific images and encoders that were otherwise pre-trained on natural images.
The results highlight the importance of using domain-specific training data and focusing on learning deep structural patterns for best pre-training performance.
|
