Home

Awesome

DoNet: Deep De-overlapping Network for Cytology Instance Segmentation [CVPR2023]

This is the official pytorch implementation of DoNet, please refer the paper and demo for more details.

Introduction

Abstract

Cell instance segmentation in cytology images has significant importance for biology analysis and cancer screening, while remains challenging due to 1) the extensive overlapping translucent cell clusters that cause the ambigu- ous boundaries, and 2) the confusion of mimics and debris as nuclei. In this work, we proposed a De-overlapping Network (DoNet) in a decompose-and-recombined strategy. A Dual-path Region Segmentation Module (DRM) explicitly decomposes the cell clusters into intersection and complement regions, followed by a Semantic Consistency-guided Recombination Module (CRM) for integration. To further introduce the containment relationship of the nucleus in the cytoplasm, we design a Mask-guided Region Proposal Strategy (MRP) that integrates the cell attention maps for inner-cell instance prediction. We validate the proposed approach on ISBI2014 and CPS datasets. Experiments show that our proposed DoNet significantly outperforms other state-of-the-art (SOTA) cell instance seg- mentation methods.

Visualization

Results

ISBI2014 dataset

MethodsmAPDiceF1AJI
Mask R-CNN59.0991.1592.5477.07
Cascade R-CNN62.4591.2992.5177.91
Mask Scoring R-CNN63.5691.2891.8775.14
HTC59.6291.3988.0875.00
Occlusion R-CNN62.3591.7593.1878.64
Xiao et al.57.3491.7092.7578.29
DoNet64.0292.1393.2379.05

CPS dataset

MethodsmAPDiceF1AJI
Mask R-CNN48.28 $\pm$ 3.1089.32 $\pm$ 0.5085.07 $\pm$ 2.0169.20 $\pm$ 2.27
Cascade R-CNN47.87 $\pm$ 3.2789.24 $\pm$ 0.4483.33 $\pm$ 1.6568.86 $\pm$ 3.55
Mask Scoring R-CNN48.38 $\pm$ 3.1389.39 $\pm$ 0.2482.98 $\pm$ 1.8667.45 $\pm$ 2.45
HTC47.60 $\pm$ 3.5689.08 $\pm$ 0.5181.30 $\pm$ 2.5666.35 $\pm$ 2.84
Occlusion R-CNN48.14 $\pm$ 2.8489.08 $\pm$ 0.2885.69 $\pm$ 2.2869.51 $\pm$ 2.45
Xiao et al.48.53 $\pm$ 2.8589.29 $\pm$ 0.2485.46 $\pm$ 2.6069.37 $\pm$ 2.88
DoNet49.43 $\pm$ 3.8389.54 $\pm$ 0.2585.51 $\pm$ 2.3370.08 $\pm$ 2.84
DoNet w/ Aug.49.65 $\pm$ 3.5289.50 $\pm$ 0.3886.30 $\pm$ 2.0170.56 $\pm$ 2.34

Dataset

Download official ISBI 2014 dataset from Overlapping Cervical Cytology Image Segmentation Challenge.

You can also download our preprocessed ISBI 2014 datasets (COCO format annotations, amodal instance GT generation).

Installation

See INSTALL.md.

QuickStart

Inference and Visualization with Pre-trained Models (on ISBI)

Use customized_tools/inference.py

python inference.py 
    --annTestFile <path to the test json file> 
    --imgTestFile <path to the test image file> 
    --configFile <path to the config file> 
    --outputDir <path to the output directory>
    --weightsFile <name of the weights file>

Download ISBI pre-trained model

Evaluation with Pre-trained Models (on ISBI)

Use customized_tools/evaluate.py The evaluation code rely on a modified pycoco package that provides a new function iouIntUni to compute intersection over union between masks, return iou, intersection, union together. For installation of the modified pycoco package, please refer to https://github.com/Amandaynzhou/MMT-PSM

python customized_tools/evaluate.py 
    --annTestFile <path to the test json file> 
    --imgTestFile <path to the test image file> 
    --configFile <path to the config file> 
    --outputDir <path to the output directory>
    --weightsFile <name of the weights file>

Training on your own dataset

Use customized_tools/preprocess.py to process your dataset.

python customized_tools/preprocess.py 
    --json_path <path to the original json file> 
    --output_json_path <path to the output json file>

Use customized_tools/train.py to train your model.

python train.py 
    --annTrainFile <path to the train json file>
    --imgTrainFile <path to the train image file>
    --annValFile <path to the validation json file> 
    --imgValFile <path to the validation image file> 
    --configFile <path to the config file> 
    --outputDir <path to the output directory>

Config Files

<!-- table -->

(Please refer to Table 3 in the paper)

modelconfig file
BaseCOCO-InstanceSegmentation/mask_rcnn_R_50_FPN_3x.yaml
Base + DRMCOCO-InstanceSegmentation/mask_triple_branch_rcnn_R_50_FPN_3x_DRM.yaml
Base + DRM + CRMCOCO-InstanceSegmentation/mask_triple_branch_rcnn_R_50_FPN_3x_DRM_CRM.yaml
Base + DRM + CRM + MRPCOCO-InstanceSegmentation/mask_triple_branch_rcnn_R_50_FPN_3x_DRM_CRM_MRP.yaml

License

Detectron2 is released under the Apache 2.0 license.

Citation

@InProceedings{Jiang_2023_CVPR,
    author    = {Jiang, Hao and Zhang, Rushan and Zhou, Yanning and Wang, Yumeng and Chen, Hao},
    title     = {DoNet: Deep De-Overlapping Network for Cytology Instance Segmentation},
    booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
    month     = {June},
    year      = {2023},
    pages     = {15641-15650}
}

Acknowledgment

The code of DoNet is built on detectron2 and ORCNN, many thanks for the Third Party Libs.

Question

Feel free to email us if you have any questions:

Rushan Zhang(rzhangbq@connect.ust.hk), Hao Jiang(hjiangaz@cse.ust.hk)