Master thesis : Image Segmentation of Zebrafish Images

Abstract

The Biomedaqu project's goal is to find solutions for skeletal disorders. They have been taking advantage of the transparent body of zebrafish as a way to study the development of the skeleton under various conditions. However, their work is slowed down by the time-consuming task of annotating images. This project aims to automate the annotation process using image segmentation. However, this project will run into the problem of class imbalance as the many features to extract from zebrafish are extremely small. To counter this problem, two methods will be used. Cost-level methods will be used to study the effect of various loss functions on the performances of different models, while a data-level method will be used in order to reduce the class imbalance in the data. Even though not all methods were proven to be successful in dealing with class imbalance, this work provides valuable insight and findings that can guide future research in this area.