Domain Adaptive Mitochondria Segmentation via Enforcing Inter-Section Consistency [paper]

Accepted by MICCAI-2022

Wei Huang, Xiaoyu Liu, Zhen Cheng, Yueyi Zhang, and Zhiwei Xiong*

University of Science and Technology of China (USTC), Hefei, China

Institute of Artificial Intelligence, Hefei Comprehensive National Science Center, Hefei, China

*Corresponding Author

Abstract

Deep learning-based methods for mitochondria segmentation require sufficient annotations on Electron Microscopy (EM) volumes, which are often expensive and time-consuming to collect. Recently, Unsupervised Domain Adaptation (UDA) has been proposed to avoid annotating on target EM volumes by exploiting annotated source EM volumes. However, existing UDA methods for mitochondria segmentation only address the intra-section gap between source and target volumes but ignore the inter-section gap between them, which restricts the generalization capability of the learned model on target volumes. In this paper, for the first time, we propose a domain adaptive mitochondria segmentation method via enforcing inter-section consistency. The key idea is to learn an inter-section residual on the segmentation results of adjacent sections using a CNN. The inter-section residuals predicted from source and target volumes are then aligned via adversarial learning. Meanwhile, guided by the learned inter-section residual, we can generate pseudo labels to supervise the segmentation of adjacent sections inside the target volume, which further enforces inter-section consistency. Extensive experiments demonstrate the superiority of our proposed method on four representative and diverse EM datasets. Code is available at https://github.com/weih527/DA-ISC.

Framework and Network Architecture

Environment

This code was tested with Pytorch 1.0.1 (later versions may work), CUDA 9.0, Python 3.7.4 and Ubuntu 16.04. It is worth mentioning that, besides some commonly used image processing packages.

If you have a Docker environment, we strongly recommend you to pull our image as follows,

docker pull registry.cn-hangzhou.aliyuncs.com/renwu527/auto-emseg:v5.4

or

docker pull renwu527/auto-emseg:v5.4

Besides, we need to instanll some python packages manually:

pip install albumentations
pip uninstall opencv-python  # remove the old version
pip install opencv-python

The entire installed python packages can be found in 'requirements.txt'

Datasets

Data Properties

Datasets	VNC III	Lucchi	MitoEM-R	MitoEM-H
Organism	Drosophila	Mouse	Rat	Human
Tissue	Ventral nerve cord	Hippocampus	Cortex	Cortex
Device	ssTEM	FIB-SEM	mbSEM	mbSEM
Resolution	50x5x5 nm	5x5x5 nm	30x8x8 nm	30x8x8 nm
Training set	20x1024x1024	165x768x1024	400x4096x4096	400x4096x4096
Test set	None	165x768x1024	100x4096x4096	100x4096x4096
Website	GitHub	EPFL	MitoEM	MitoEM

You can download our processed data directly from GoogleDrive or BaiduYun (Access code: weih). However, because the MitoEM dataset is too large (>10GB), we cannot put it in our cloud storage. It is recommended to download it from the official website.

Data Tree

|--./data
|   |--Lucchi
|   |   |--testing
|   |   |--testing_groundtruth
|   |   |--training
|   |   |--training_groundtruth
|   |--Mito
|   |   |--human
|   |   |   |--testing.hdf
|   |   |   |--testing_groundtruth.hdf
|   |   |   |--training.hdf
|   |   |   |--training_groundtruth.hdf
|   |   |--rat
|   |   |   |--testing.hdf
|   |   |   |--testing_groundtruth.hdf
|   |   |   |--training.hdf
|   |   |   |--training_groundtruth.hdf
|   |--VNC3
|   |   |--training
|   |   |--training_groundtruth

Training

We train our method on one NVIDIA Tianxp GPU. Our training log files can be found in './logs'.

VNC III --> Lucchi (Subset1)

cd scripts
python main.py -c vnc2lucchi1

VNC III --> Lucchi (Subset2)

cd scripts
python main.py -c vnc2lucchi2

MitoEM-R --> MitoEM-H

cd scripts
python main_mito.py -c mitor2h

MitoEM-H --> MitoEM-R

cd scripts
python main_mito.py -c mitoh2r

Inference

We test our trained model on one NVIDIA Tianxp GPU.

We store our trained models at GoogleDrive or BaiduYun (Access code: weih)

VNC III --> Lucchi (Subset1)

cd scripts
python inference.py -c vnc2lucchi1 -mn vnc2lucchi1 -sw

Print

cfg_file: vnc2lucchi1.yaml
out_path: ../inference/vnc2lucchi1
Begin inference...
Prediction time (s): 138.40105080604553
Measure on mAP, F1, MCC, and IoU...
mAP=0.8948, F1=0.8129, MCC=0.8053, IoU=0.6865
Measurement time (s): 917.6475455760956
Done

VNC III --> Lucchi (Subset2)

cd scripts
python inference.py -c vnc2lucchi2 -mn vnc2lucchi2 -sw

Print

cfg_file: vnc2lucchi2.yaml
out_path: ../inference/vnc2lucchi2
Begin inference...
Prediction time (s): 144.69077563285828
Measure on mAP, F1, MCC, and IoU...
mAP=0.9244, F1=0.8518, MCC=0.8448, IoU=0.7431
Measurement time (s): 912.5876989364624
Done

MitoEM-R --> MitoEM-H

It needs large memory for quantitative measurement (>100GB)

cd scripts
python inference_mito.py -c mitor2h -mn mitor2h -sw

Print

cfg_file: mitor2h.yaml
out_path: ../inference/mitor2h
Load ../data/Mito/human/testing.hdf
raw shape: (100, 4096, 4096)
padded raw shape: (100, 4608, 4608)
iters: 6336
Load ../data/Mito/human/testing_groundtruth.hdf
Begin inference...
the number of sub-volume: 6336
Prediction time (s): 1438.0159723758698
Measure on mAP, F1, MCC, and IoU...
mAP=0.9256, F1=0.8557, MCC=0.8495, IoU=0.7479
Measurement time (s): 4127.024113416672
Done

MitoEM-H --> MitoEM-R

It needs large memory for quantitative measurement (>100GB)

cd scripts
python inference_mito.py -c mitoh2r -mn mitoh2r -sw

Print

cfg_file: mitoh2r.yaml
out_path: ../inference/mitoh2r
Load ../data/Mito/rat/testing.hdf
raw shape: (100, 4096, 4096)
padded raw shape: (100, 4608, 4608)
iters: 6336
Load ../data/Mito/rat/testing_groundtruth.hdf
Begin inference...
the number of sub-volume: 6336
Prediction time (s): 1441.9779460430145
Measure on mAP, F1, MCC, and IoU...
mAP=0.9682, F1=0.8851, MCC=0.8829, IoU=0.7941
Measurement time (s): 4129.04722571373
Done

Visual Results

Contact

If you have any problem with the released code, please do not hesitate to contact me by email ([email protected]).