Department of Computer Engineering
DEEP LEARNING APPROACHES FOR CELL INSTANCE SEGMENTATION
Melih Utku Şahin
(Supervisor: Assoc. Prof. Dr. Çiğdem Gündüz Demir)
Computer Engineering Department
Semantic segmentation is a task of image classification. While in classification, the main purpose is to find the class that whole image belong to, our purpose in segmentation is to classify each pixel within the image to divide it into segments. It has lots of different applications varying from autonomous driving to medical image analysis. During the past years with the advancements in Convolutional Neural Networks, deep learning became the driving methodology behind the semantic segmentation task. With the recent improvements in semantic segmentation based methodologies, fields that highly rely on significant computational works had great improvements. Among segmentation tasks, medical image analysis is hard to interpret and computationally heavy due to very large image resolutions. Although we have large images, we have very limited datasets in terms of image quantity. Also, dataset images are highly imbalanced since most of the data samples are consisting of background pixels and positive pixel volume is significantly small. In this work, we will be focusing on semantic segmentation task for cell instance segmentation. We will be using various deep network architectures, loss functions and data preprocessing techniques. Our main goal is to overcome the negative effects of class imbalance, extract cell bodies and boundaries successfully and train our system effectively with the perks of them.
DATE: 19 April 2021, Monday @ 16:00