Label-efficient segmentation of organoid culture data using diffusion models

Wiers, Jesse (2023) Label-efficient segmentation of organoid culture data using diffusion models. Bachelor's Thesis, Artificial Intelligence.

Preview	Text Bachelor_Thesis_Final_JesseWiers_s3998576.pdf Download (4MB) | Preview
	Text toestemming.pdf Restricted to Registered users only Download (130kB)

Abstract

As the volume of biomedical data continues to expand at a rapid pace, the potential of extracting valuable insights from this data through deep learning models is also increasing. However, this process typically necessitates labeled data, which has traditionally been manually annotated. This manual approach is associated with various constraints, including time, finan- cial resources, and expertise, and it can also be prone to errors due to fatigue. The objective of this study is to utilize diffusion models, specifically diffusion denoising probabilistic models (DDPMs), for the segmentation of organoid culture data. Two different methods are employed using DPPMs. Firstly, segmentation will be carried out by utilizing feature maps of DDPMs that have been trained to generate samples of organoid culture data. These feature maps will be combined with an ensemble of multi-layer perceptrons. Secondly, DDPMs will be trained to directly generate segmentation maps for organoid culture data. The methods were evaluated on the MIoU, Dice and HD95 score on a maximum of 42.348 images. On 100% of the data, the representation approach (MIoU=0.92, Dice=0.96, HD95=35) outperformed the direct segmenta- tion approach (MIoU=0.62, Dice=0.71, HD95=62) for all metrics. The representation approach also proved to be suitable for label-efficient segmentation since the aforementioned performance for the representation approach is achieved with as little as 20 labelled images in the training pipeline

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:	23 Aug 2023 14:13
Last Modified:	23 Aug 2023 14:13
URI:	https://fse.studenttheses.ub.rug.nl/id/eprint/31251

Actions (login required)

View Item