Towards Automatic 3D Shape Instantiation for Deployed Stent Grafts: 2D Multiple-class and Class-imbalance Marker Segmentation with Equally-weighted Focal U-Net

摘要

Small object segmentation is a common task in medical image analysis.Traditional feature-based methods require human intervention while methodsbased on deep learning train the neural network automatically. However, it isstill error prone when applying deep learning methods for small objects. Inthis paper, Focal FCN was proposed for small object segmentation with limitedtraining data. Firstly, Fully-weighted FCN was proposed to apply aninitialization for Focal FCN by adding weights to the background and foregroundloss. Secondly, focal loss was applied to make the training focus onwrongly-classified pixels and hence achieve good performance on small objectsegmentation. Comparisons between FCN, Weighted FCN, Fully-weighted FCN andFocal FCN were tested on customized stent graft marker segmentation.