medical image segmentation: a survey medical image segmentation
Three architectures achieve good performance:
structure: input-conv-deconv-segmentation
Implementation: ultrasound-nerve-segmentation
propose skip-connect from encoder to decoder.
1. propose dice-coeff loss which is specially designed for medical image segmentation. 2. illustrate the idea of 3D convolution on a images volume.
Other useful paper:
have a similar skip-connect structure
, symmetric upsampling. not as powerful as u-net.
<DHSnet: Deep Hierarchical Saliency Network for Saliency Object Detection> similar to medical image (predict single probability map)
structure: input-conv-upsampling-refinement
Implementation: medical-image-segmentation
<Deeplab: Semantic Image Segmentation with Deep Convolutional Nets, Atrous Convolution, and Fully Connected CRFs> achieves best performance on PASCAL VOC segmentation challenge
similar to medical image (predict single probability map)
<DCAN: Deep Contour-Aware Networks for Accurate Gland Segmentation> learning to predict coutour map and segmentation map.
SDS first propose the task:
Implementation: rpn_drn_new
defeat segmentation-only encoder-decoder network on COCO segmentation challenge
Miscellaneous
Dilated convolution helps reduce computation