- Abstract: Screening for increased risk of pregnancy complications could be possible with fully automated placental segmentation in 3D ultrasound (3D-US). Fully convolutional neural networks (fCNN) have previously obtained good segmentation performance of the first trimester placenta and appears to predict fetal growth restriction better than manual segmentation methods. The goal of this study is to adjust fCNN architecture parameters to investigate their impact on performance and ultimately to produce a more accurate segmentation. 2,393 first trimester 3D-US volumes with ‘ground-truth’ segmentation obtained using a semi-automated technique were used. An open source package (OxNNet) was used to train end-to-end six fully convolutional neural networks with different loss functions, addition of batch normalisation and with different numbers of features. A small increase in performance of placental segmentation in terms of Dice similarity coefficient (DSC) (0.835 vs 0.825) was observed. Doubling the feature map gave a minor improvement in DSC (0.01). Use of batch normalisation increased the speed of training as expected. The Dice-based loss gave poorer performance in general. Convolution with no padding produced better segmentation than using padding. The subjective case quality assessment score was shown to correlate with the DSC (r = -0.28 (p < 0.05)). A faster, less-memory intensive fCNN architecture can provide a similar segmentation performance moving the use of this tool for clinical screening a step closer.
- Keywords: Placenta, 3D Ultrasound, Pregnancy
- Author Affiliation: Nuffield Department of Women’s & Reproductive Health, University of Oxford, Level 3, Women’s Centre, John Radcliffe Hospital, Oxford OX3 9DU., School of Women’s and Children’s Health, University of New South Wales, Randwick, NSW, Australia., Harris Birthright Research Centre of Fetal Medicine, King’s College Hospital, UK., Fetal Medicine Unit. Hospiten Group. Tenerife. Canary Islands. Spain., Fetal Medicine Unit, The Women’s Centre, John Radcliffe Hospital Oxford, Department of Obstetrics and Gynaecology, Wexham Park Hospital, Slough