hide
Language(s):
eng - English
Dates:
2018-05-022018
Publication Status:
Published in print
Pages:
86 p.
Publishing info:
Saarbrücken : Universität des Saarlandes
Table of Contents:
Automatising the process of semantic segmentation of anatomical structures in medical data is an active field of research in Deep Learning. Recent advances in Convo-lutional Neural Networks for various image related tasks have inspired researchers to adopt CNN based methods for semantic segmentation of bio-medical images. In this thesis we explore different CNN architectures and frameworks with the aim to build an optimal model for automatic segmentation of CT images with Bone-Implant systems. For this purpose, we evaluate three CNN architectures - SegNet, UNet and 3D-UNet and address the challenges encountered during the network implementation on the available data. These challenges include scarcity of labelled data, class imbalance and hyperparameter selection and we try to overcome them by using techniques such as data augmentation, weighted loss functions and cyclical learning rate. We show that data augmentation when combined with cyclical learning rate method using UNet not only trains the model in less time and but also achieves better accuracy for minority classes.
We also investigate into two deep learning frameworks - TensorFlow and Caffe2 using Python language for training and inference of CNNs. Caffe2 is relatively new framework, created to overcome the drawbacks of Caffe. Out of the two networks TensorFlow is the preferred choice for framework as it provides user convenience in all aspects like installation, development and deployment.
Rev. Type:
-
Identifiers:
BibTex Citekey: GroverMaster2018
Degree:
Master
