mmsegmentation/configs/unet

• UNet
• • Introduction
  • Abstract
  • Citation
  • Results and models
  • • Cityscapes
    • DRIVE
    • STARE
    • CHASE_DB1
    • HRF

UNet

U-Net: Convolutional Networks for Biomedical Image Segmentation

Introduction

Official Repo

Code Snippet

Abstract

There is large consent that successful training of deep networks requires many thousand annotated training samples. In this paper, we present a network and training strategy that relies on the strong use of data augmentation to use the available annotated samples more efficiently. The architecture consists of a contracting path to capture context and a symmetric expanding path that enables precise localization. We show that such a network can be trained end-to-end from very few images and outperforms the prior best method (a sliding-window convolutional network) on the ISBI challenge for segmentation of neuronal structures in electron microscopic stacks. Using the same network trained on transmitted light microscopy images (phase contrast and DIC) we won the ISBI cell tracking challenge 2015 in these categories by a large margin. Moreover, the network is fast. Segmentation of a 512x512 image takes less than a second on a recent GPU. The full implementation (based on Caffe) and the trained networks are available at this http URL.

Citation

@inproceedings{ronneberger2015u,
  title={U-net: Convolutional networks for biomedical image segmentation},
  author={Ronneberger, Olaf and Fischer, Philipp and Brox, Thomas},
  booktitle={International Conference on Medical image computing and computer-assisted intervention},
  pages={234--241},
  year={2015},
  organization={Springer}
}

Results and models

Cityscapes

Method	Backbone	Loss	Crop Size	Lr schd	Mem (GB)	Inf time (fps)	mIoU	mIoU(ms+flip)	config	download
UNet + FCN	UNet-S5-D16	Cross Entropy	512x1024	160000	17.91	3.05	69.10	71.05	config	model | log

DRIVE

Method	Backbone	Loss	Image Size	Crop Size	Stride	Lr schd	Mem (GB)	Inf time (fps)	mDice	Dice	config	download
UNet + FCN	UNet-S5-D16	Cross Entropy	584x565	64x64	42x42	40000	0.680	-	88.38	78.67	config	model | log
UNet + FCN	UNet-S5-D16	Cross Entropy + Dice	584x565	64x64	42x42	40000	0.582	-	88.71	79.32	config	model | log
UNet + PSPNet	UNet-S5-D16	Cross Entropy	584x565	64x64	42x42	40000	0.599	-	88.35	78.62	config	model | log
UNet + PSPNet	UNet-S5-D16	Cross Entropy + Dice	584x565	64x64	42x42	40000	0.585	-	88.76	79.42	config	model | log
UNet + DeepLabV3	UNet-S5-D16	Cross Entropy	584x565	64x64	42x42	40000	0.596	-	88.38	78.69	config	model | log
UNet + DeepLabV3	UNet-S5-D16	Cross Entropy + Dice	584x565	64x64	42x42	40000	0.582	-	88.84	79.56	config	model | log

STARE

Method	Backbone	Loss	Image Size	Crop Size	Stride	Lr schd	Mem (GB)	Inf time (fps)	mDice	Dice	config	download
UNet + FCN	UNet-S5-D16	Cross Entropy	605x700	128x128	85x85	40000	0.968	-	89.78	81.02	config	model | log
UNet + FCN	UNet-S5-D16	Cross Entropy + Dice	605x700	128x128	85x85	40000	0.986	-	90.65	82.70	config	model | log
UNet + PSPNet	UNet-S5-D16	Cross Entropy	605x700	128x128	85x85	40000	0.982	-	89.89	81.22	config	model | log
UNet + PSPNet	UNet-S5-D16	Cross Entropy + Dice	605x700	128x128	85x85	40000	1.028	-	90.72	82.84	config	model | log
UNet + DeepLabV3	UNet-S5-D16	Cross Entropy	605x700	128x128	85x85	40000	0.999	-	89.73	80.93	config	model | log
UNet + DeepLabV3	UNet-S5-D16	Cross Entropy + Dice	605x700	128x128	85x85	40000	1.010	-	90.65	82.71	config	model | log

CHASE_DB1

Method	Backbone	Loss	Image Size	Crop Size	Stride	Lr schd	Mem (GB)	Inf time (fps)	mDice	Dice	config	download
UNet + FCN	UNet-S5-D16	Cross Entropy	960x999	128x128	85x85	40000	0.968	-	89.46	80.24	config	model | log
UNet + FCN	UNet-S5-D16	Cross Entropy + Dice	960x999	128x128	85x85	40000	0.986	-	89.52	80.40	config	model | log
UNet + PSPNet	UNet-S5-D16	Cross Entropy	960x999	128x128	85x85	40000	0.982	-	89.52	80.36	config	model | log
UNet + PSPNet	UNet-S5-D16	Cross Entropy + Dice	960x999	128x128	85x85	40000	1.028	-	89.45	80.28	config	model | log
UNet + DeepLabV3	UNet-S5-D16	Cross Entropy	960x999	128x128	85x85	40000	0.999	-	89.57	80.47	config	model | log
UNet + DeepLabV3	UNet-S5-D16	Cross Entropy + Dice	960x999	128x128	85x85	40000	1.010	-	89.49	80.37	config	model | log

HRF

Method	Backbone	Loss	Image Size	Crop Size	Stride	Lr schd	Mem (GB)	Inf time (fps)	mDice	Dice	config	download
UNet + FCN	UNet-S5-D16	Cross Entropy	2336x3504	256x256	170x170	40000	2.525	-	88.92	79.45	config	model | log
UNet + FCN	UNet-S5-D16	Cross Entropy + Dice	2336x3504	256x256	170x170	40000	2.623	-	89.64	80.87	config	model | log
UNet + PSPNet	UNet-S5-D16	Cross Entropy	2336x3504	256x256	170x170	40000	2.588	-	89.24	80.07	config	model | log
UNet + PSPNet	UNet-S5-D16	Cross Entropy + Dice	2336x3504	256x256	170x170	40000	2.798	-	89.69	80.96	config	model | log
UNet + DeepLabV3	UNet-S5-D16	Cross Entropy	2336x3504	256x256	170x170	40000	2.604	-	89.32	80.21	config	model | log
UNet + DeepLabV3	UNet-S5-D16	Cross Entropy + Dice	2336x3504	256x256	170x170	40000	2.607	-	89.56	80.71	config	model | log

Note:

• In DRIVE, STARE, CHASE_DB1, and HRF dataset, mDice is mean dice of background and vessel, while Dice is dice metric of vessel(foreground) only.