TY  - THES
U1  - Master Thesis
A1  - Paskali, Filip
T1  - Automatic prostate segmentation in transrectal ultrasound images using modified V-net convolutional neural network
N2  - Prostate segmentation is an essential part of brachytherapy treatment planning, in order to perform the procedure with required accuracy. Nowadays, segmentation of the prostate is still carried out manually during the planning steps, therefore it is a process that can be tedious, time-consuming and prone to inter-observer error. Much effort has been made in development of an computer-based algorithm that can perform prostate segmentation automatically, but only with appearance of deep learning methods, more promising algorithms emerged. So far, convolutional neural networks demonstrated excellent results in fully automatic prostate segmentation. Development of such an algorithm and training an efficient deep learning model is a challenging task, and requires a lot of optimizations. The objective of this study is development and evaluation of an algorithm for image processing based on deep learning methods that can perform fully automatic segmentation of the prostate gland in transrectal ultrasound images. Additionally, we made an overview of the development process, along with challenges and their solutions and demonstrated an algorithm implemented using Python and Tensorflow library, consisted of preprocessing, augmentation, training and validation, postprocessing and validation steps, which is able to successfully carry out fully automatic prostate segmentation with expert level of accuracy. Finally, we presented our implementation of fully convolutional neural network model and results that are encouraging to continue with model improvements and potential clinical application.
KW  - Deep-learning
KW  - Segmentation
KW  - Prostate
KW  - Image-analysis
KW  - Ultrasound
Y2  - 2020
U6  - https://nbn-resolving.org/urn:nbn:de:bsz:fn1-opus4-73433
UN  - https://nbn-resolving.org/urn:nbn:de:bsz:fn1-opus4-73433
UR  - https://github.com/fpaskali/trus-segmentation
N1  - Der Code zur Masterthesis ist unter der genannten url verfügbar.
SP  - 61
S1  - 61
ER  -