Original title: Denoise Pre-Training For Segmentation Neural Networks
Authors: Kolarik, Martin
Document type: Papers
Language: eng
Publisher: Vysoké učení technické v Brně, Fakulta elektrotechniky a komunikačních technologií
Abstract: This paper proposes a method for pre-training segmentation neural networks on small datasets using unlabelled training data with added noise. The pre-training process helps the network with initial better weights settings for the training itself and also augments the training dataset when dealing with small labelled datasets especially in medical imaging. The experiment comparing results of pre-trained and not pre-trained networks on MRI brain segmentation task has shown that the denoise pre-training helps the network with faster training convergence without overfitting and achieving better results in all compared metrics even on very small datasets.
Keywords: deep learning; denoising; neural network; pre-training; segmentation
Host item entry: Proceedings of the 25st Conference STUDENT EEICT 2019, ISBN 978-80-214-5735-5

Institution: Brno University of Technology (web)
Document availability information: Fulltext is available in the Brno University of Technology Digital Library.
Original record: http://hdl.handle.net/11012/186770

Permalink: http://www.nusl.cz/ntk/nusl-414672


The record appears in these collections:
Universities and colleges > Public universities > Brno University of Technology
Conference materials > Papers