PET Tumor Segmentation Using a Deep Residual CNN with Dilated Convolutions

Van der Veen, Werner (2018) PET Tumor Segmentation Using a Deep Residual CNN with Dilated Convolutions. Bachelor's Thesis, Artificial Intelligence.

Preview	Text Bachelors_Thesis_Werner_Final.pdf Download (447kB) | Preview
	Text toestemming.pdf Restricted to Registered users only Download (94kB)

Abstract

This research examines the effectiveness of using deep learning for medical image segmentation. Here, tumors are segmented from a dataset of three-dimensional PET-scans. A deep convolutional neural network is trained on this dataset, and approaches from natural image segmentation are tested, particularly residual connections and dilated convolution. Its segmentation performance is compared to that of a simple thresholding algorithm that segments based on voxel intensity values. Using a weighted Jaccard index metric and loss and a positive predictive value and sensitivity metric, the neural network appears to outperform the simple thresholding algorithm slightly but consistently. However, additional varied research is needed to corroborate the findings and the advantages of deep learning for medical image segmentation in general.

Item Type:	Thesis (Bachelor's Thesis)
Supervisor name:	Wiering, M.A.
Degree programme:	Artificial Intelligence
Thesis type:	Bachelor's Thesis
Language:	English
Date Deposited:	30 Jul 2018
Last Modified:	31 Jul 2018 12:20
URI:	https://fse.studenttheses.ub.rug.nl/id/eprint/18148

Actions (login required)

View Item