Semi-supervised contrastive learning for organoid microscopy image segmentation

Zwijghuizen, Julian (2023) Semi-supervised contrastive learning for organoid microscopy image segmentation. Bachelor's Thesis, Artificial Intelligence.

Preview	Text bachelor_thesis_Julian_Zwijghuizen.pdf Download (2MB) | Preview
	Text toestemming.pdf Restricted to Registered users only Download (130kB)

Abstract

This project employs a semi-supervised learning approach for segmenting organoid microscopy images, involving two distinct stages: pre-training and fine-tuning. The pre-training stage can be further divided into a global unsupervised contrastive learning stage and a local supervised contrastive learning stage. The objective is to investigate whether the semi-supervised approach outperforms the supervised approach. To evaluate this, the models are trained on varying amounts of data during the pre-training stage to determine the minimum quantity required to develop a model that outperforms the supervised learning approach. Additionally, the study examines whether two different loss functions (SSIM loss and SSIM-MAE loss) positively contribute to the segmentation performance when being used in the fine-tuning stage. Finally, the effect of freezing (vs not freezing) the U-Net encoder of the global stage when training on the local stage is examined in the context of segmentation performance. Results showed that the local stage of the semi-supervised learning approach has a more positive impact on the F1-scores as more data is used compared to the global stage, with F1-scores around 0.9. The SSIM(-MAE) is a better choice in terms of the organoids' coherent structure and the frozen models outperform the frozen models in capturing the larger organoids.

Item Type:	Thesis (Bachelor's Thesis)
Supervisor name:	Schomaker, L.R.B. and Haja, A.
Degree programme:	Artificial Intelligence
Thesis type:	Bachelor's Thesis
Language:	English
Date Deposited:	02 Aug 2023 07:54
Last Modified:	02 Aug 2023 07:54
URI:	https://fse.studenttheses.ub.rug.nl/id/eprint/30875

Actions (login required)

View Item