U-Net: Convolutional Networks for Biomedical Image Segmentation

18 May 2015  ·  Olaf Ronneberger, Philipp Fischer, Thomas Brox ·

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 http://lmb.informatik.uni-freiburg.de/people/ronneber/u-net .

PDF Abstract

Results from the Paper


Task Dataset Model Metric Name Metric Value Global Rank Result Benchmark
Lesion Segmentation Anatomical Tracings of Lesions After Stroke (ATLAS) U-Net Dice 0.4606 # 4
IoU 0.3447 # 3
Precision 0.5994 # 2
Recall 0.4449 # 3
Retinal Vessel Segmentation CHASE_DB1 U-Net AUC 0.9772 # 11
Pancreas Segmentation CT-150 U-Net Dice Score 0.814 # 1
Precision 0.848 # 2
Recall 0.806 # 2
Medical Image Segmentation CVC-ClinicDB U-Net mean Dice 0.8230 # 33
Dichotomous Image Segmentation DIS-TE1 UNet max F-Measure 0.625 # 16
weighted F-measure 0.514 # 16
MAE 0.106 # 18
S-Measure 0.716 # 15
E-measure 0.750 # 17
HCE 233 # 9
Dichotomous Image Segmentation DIS-TE2 UNet max F-Measure 0.703 # 15
weighted F-measure 0.597 # 16
MAE 0.107 # 17
S-Measure 0.755 # 14
HCE 474 # 6
Dichotomous Image Segmentation DIS-TE3 UNet max F-Measure 0.748 # 12
weighted F-measure 0.644 # 15
MAE 0.098 # 16
HCE 883 # 5
Dichotomous Image Segmentation DIS-TE4 UNet max F-Measure 0.759 # 9
weighted F-measure 0.659 # 11
MAE 0.102 # 11
E-measure 0.821 # 15
HCE 3218 # 5
Dichotomous Image Segmentation DIS-VD UNet max F-Measure 0.692 # 12
weighted F-measure 0.586 # 14
MAE 0.113 # 16
S-Measure 0.745 # 11
E-measure 0.785 # 16
HCE 1337 # 6
Semantic Segmentation Event-based Segmentation Dataset U-Net mIoU 64.7 # 5
Medical Image Segmentation ISBI 2012 EM Segmentation U-Net Warping Error 0.000353 # 1
Skin Cancer Segmentation Kaggle Skin Lesion Segmentation U-Net F1 score 0.8682 # 3
AUC 0.9371 # 3
Semantic Segmentation Kvasir-Instrument UNet mIoU 0.8578 # 1
Medical Image Segmentation Kvasir-SEG U-Net Average MAE 0.055 # 12
mean Dice 0.8180 # 43
S-Measure 0.858 # 10
max E-Measure 0.893 # 10
Thermal Image Segmentation MFN Dataset UNet mIOU 45.1 # 41
Thermal Image Segmentation PST900 UNet mIoU 52.8 # 16
Medical Image Segmentation RITE U-Net Dice 55.24 # 2
Jaccard Index 31.11 # 3
Retinal Vessel Segmentation ROSE-1 DVC U-Net Dice Score 66.05 # 2
Retinal Vessel Segmentation ROSE-1 SVC U-Net Dice Score 71.16 # 5
Retinal Vessel Segmentation ROSE-1 SVC-DVC U-Net Dice Score 70.12 # 5
Retinal Vessel Segmentation ROSE-2 U-Net Dice Score 65.64 # 5
Semantic Segmentation SELMA UNet mIoU 36.2 # 7
Semantic Segmentation SkyScapes-Dense U-Net Mean IoU 14.15 # 7
Electron Microscopy Image Segmentation SNEMI3D U-Net AUC 0.8676 # 2
Semantic Segmentation Stanford2D3D Panoramic U-Net mIoU 35.9% # 25
Optic Disc Segmentation STARE U-Net AUC 0.8312 # 1
Retinal Vessel Segmentation STARE U-Net F1 score 0.8373 # 4
AUC 0.7783 # 5
Cell Segmentation STARE U-Net AUC 0.7756 # 1
Video Polyp Segmentation STARE UNet AUC 0.459 # 1
Colorectal Gland Segmentation: STARE U-Net (e) AUC 0.827 # 2
Colorectal Gland Segmentation: STARE U-Net AUC 0.835 # 1
Semantic Segmentation STARE UNet AUC 0.9158 # 1
Dichotomous Image Segmentation STARE UNet AUC 0.780 # 1
Colorectal Gland Segmentation: STARE FCN8 (e) AUC 0.796 # 3
Video Polyp Segmentation SUN-SEG-Easy UNet S measure 0.669 # 2
mean E-measure 0.677 # 2
Video Polyp Segmentation SUN-SEG-Easy (Unseen) UNet Sensitivity 0.420 # 11
Video Polyp Segmentation SUN-SEG-Hard UNet S-Measure 0.670 # 2
Dice 0.542 # 2
Video Polyp Segmentation SUN-SEG-Hard (Unseen) UNet Sensitivity 0.429 # 11
Pancreas Segmentation TCIA Pancreas-CT Dataset U-Net Dice Score 0.82 # 3
Semantic Segmentation Trans10K U-Net mIoU 29.23% # 14
GFLOPs 124.55 # 13
Lesion Segmentation University of Waterloo skin cancer database U-Net Dice score 0.836 ±0.132 # 5
Semantic Segmentation UrbanLF OCR (HRNetV2-W48) mIoU (Real) 78.60 # 5
mIoU (Syn) 79.36 # 6

Results from Other Papers


Task Dataset Model Metric Name Metric Value Rank Source Paper Compare
Colorectal Gland Segmentation: CRAG U-Net (e) Dice 0.844 # 10
Hausdorff Distance (mm) 196.9 # 5
Colorectal Gland Segmentation: CRAG FCN8 (e) Hausdorff Distance (mm) 199.5 # 4
Multi-tissue Nucleus Segmentation Kumar U-Net (e) Dice 0.758 # 18
Hausdorff Distance (mm) 47.8 # 16
Optic Disc Segmentation Drishti-GS U-Net DiceOC 0.8806 # 6
DiceOD 0.9643 # 6
mIoU 0.8487 # 5
Optic Disc Segmentation REFUGE U-Net DiceOC 0.8544 # 5
DiceOD 93.08 # 6
Retinal Vessel Segmentation DRIVE U-Net F1 score 0.8142 # 14
AUC 0.9755 # 12
Lung Nodule Segmentation LUNA U-Net F1 score 0.9658 # 3
AUC 0.9784 # 3

Methods