Brain Segmentation in Rodent MR-Images Using Convolutional Neural Networks
M.Sc. Defense by Björn Sigurdsson
Segmentation is an essential step in the analysis of medical images. Rodents are widely used in experiments and manual segmentation of rodent MR-images is time-consuming, requires intimate knowledge of rodent anatomy, and is subject to variablity between image analysts. Automatic image segmentation could provide faster and more reproducible results.
This thesis presents an analysis on the feasibility and practicality of using convolutional neural-networks (CNN) to produce a brain segmentation of rodent MR-images and how biasfield correction of images and limiting training data effect the performance of CNNs. To that end, two brain segmentation-CNNs were implemented and tested on rodent MR-images. Images segmented with a registration-based atlas (RBA) were used for comparison. One CNN was trained and tested on data processed with bias-field correction and one without, and tests were carried out with increasing amounts of training data.
The resulting segmentations were evaluated using the dice-score between automatically and manually segmented images. On average the CNNs performed, worse than the RBA, although the median dice-score of CNN segmented images was higher than the median dice score of RBA segmented images. No significant difference was found between the performance of a CNN when segmenting images with and without bias-field correction.
Increasing the number of training images improved the performance of the CNNs: training with 11 images produced significantly better results than training it with 5 images. Training with 20 images marginally improved the results over 11 training images.
Segmenting rodent MR-images using CNNs is feasible and practical. Increasing the training data improved the results of the implemented CNNs up to a point. No significant improvement was produced by adding bias-field correction to the preprocessing of the images.
Supervisors: Sune Darkner & Stefan Sommer