DDTNet: A Dense Dual-Task Network for Tumor-infiltrating Lymphocyte Detection and Segmentation in Histopathological Images of Breast Cancer

This project hosts the code for implementing simultaneous Lymphocyte Detection and Segmentation: Paper

Xiaoxuan Zhang, Xiongfeng Zhu, Kai Tang, Yinghua Zhao, Zixiao Lu, Qianjin Feng,
DDTNet: A dense dual-task network for tumor-infiltrating lymphocyte detection and segmentation in histopathological images of breast cancer,
Medical Image Analysis, Volume 78, 2022,

Implementation based on FCOS.

Installation

This DDTNet implementation is based on FCOS and maskrcnn-benchmark. Therefore the installation is the same as original FCOS and maskrcnn-benchmark.

Please check INSTALL.md for installation instructions. You may also want to see the original README.md of FCOS and README.md of maskrcnn-benchmark.

Dataset

• Download BCa-lym dataset (H&E images and original manually annotated dots) from www.andrewjanowczyk.com/use-case-4-lymphocyte-detection/

• Download Post-NAT-BRCA dataset (H&E images and original manually annotated dots) from here

• If you want to use your own dataset, the data should be prepared in the format:

project
│   README.md  
└───datasets
    └───lym
        ├──Images
        │  ├──001.png
        │  ├──002.png
        │  └...
        ├──ImageSets
        │  ├──Dataset Name1
        │  │  ├──train1.txt
        │  │  ├──test1.txt
        │  │  ├──train2.txt
        │  │  ├──test2.txt
        │  │  └...
        │  ├──Dataset Name2
        │  │  └...
        │  └...
        └──Masks
           ├──mask_001.png
           ├──edgemask_001.png
           ├──Dis_001.png       
           ├──mask_002.png
           ├──edgemask_002.png
           ├──Dis_002.png
           └...

Note:

1. Images should be H&E images (RGB images). If it is not *.png, modify lym.py.
2. *.txt should include the names of all data in the training set or test set, for example, '001', '002', '003'...
3. mask*.png, edgemask*.png and Dis*.png are mask map, contour map and distance map corresponding to input image, respectively.

Training

The following command line will train ddtnet_R_101_FPN_1x on 1 GPUs with Synchronous Stochastic Gradient Descent (SGD):

python -m torch.distributed.launch \
    --nproc_per_node=1 \
    --master_port=$((RANDOM + 10000)) \
    tools/train_net.py \
    --config-file configs/ddtnet/ddtnet_R_101_FPN_1x.yaml \
    DATALOADER.NUM_WORKERS 2 \

Inference

The inference command line:

python tools/eval_net.py \
    --config-file configs/ddtnet/ddtnet_R_101_FPN_1x.yaml \
    MODEL.WEIGHT model_final.pth \
    TEST.IMS_PER_BATCH 4    

The Inference results include a list of bounding box coordinates, a mask map and a contour map for each input image. Use Alignment.m to unify the detection and segmentation results.

Acknowledgments

We would like to thank FCOS for overall framework.