Deep learning in quantifying vascular burden from brain images
Nieminen, Tuomas Juhanpoika (2018)
2018
Sähkötekniikka
Tieto- ja sähkötekniikan tiedekunta - Faculty of Computing and Electrical Engineering
This publication is copyrighted. You may download, display and print it for Your own personal use. Commercial use is prohibited.
Hyväksymispäivämäärä
2018-05-09
Julkaisun pysyvä osoite on
https://urn.fi/URN:NBN:fi:tty-201804261565
https://urn.fi/URN:NBN:fi:tty-201804261565
Tiivistelmä
White matter hyperintensities (WMH) and lacunar infarcts are features of cerebral vessel disease. Together with cortical infarcts they are main causes of vascular dementia. Also, increased WMH volume is associated with the risk of Alzheimer’s disease. This is the reason why an accurate automatic WMH and infarct segmentation tool is highly desirable in order to improve dementia diagnosis.
In this thesis deep learning, more precisely, convolutional neural network called uResNet was used to segment WMH, lacunar infarcts and cortical infarcts from brain images. The study was done by training the network using multiple different input channel sets. Also, the amount of classes to be segmented varied. In total 21 different combinations were trained and tested including both 2D and 3D models.
The numerical and visual evaluation was performed by comparing result images to the expert annotaded images. Numerical evaluation included computation of Dice scores and correlation between the image sets. Also, for infarct detection sensitivities and amount of false positive segmentations were calculated. From the results can be deduced that proposed segmentation method is capable of accurate WMH segmentation (best achieved Dice score for WMH volumes was 0.774). However, further research is still needed in order to improve infarct segmentation results since sensitivity scores were surprisingly poor and the amount of false positive segmentations was high.
In this thesis deep learning, more precisely, convolutional neural network called uResNet was used to segment WMH, lacunar infarcts and cortical infarcts from brain images. The study was done by training the network using multiple different input channel sets. Also, the amount of classes to be segmented varied. In total 21 different combinations were trained and tested including both 2D and 3D models.
The numerical and visual evaluation was performed by comparing result images to the expert annotaded images. Numerical evaluation included computation of Dice scores and correlation between the image sets. Also, for infarct detection sensitivities and amount of false positive segmentations were calculated. From the results can be deduced that proposed segmentation method is capable of accurate WMH segmentation (best achieved Dice score for WMH volumes was 0.774). However, further research is still needed in order to improve infarct segmentation results since sensitivity scores were surprisingly poor and the amount of false positive segmentations was high.