HarDNet-DFUS: Enhancing Backbone and Decoder of HarDNet-MSEG for Diabetic Foot Ulcer Image Segmentation

1st Place Winner in MICCAI DFUC 2022!

Official PyTorch implementation of HarDNet-DFUS, contains the prediction codes for our submission to the Diabetic Foot Ulcer Segmentation Challenge 2022 (DFUC2022) at MICCAI2022.

HarDNet Family

inference on V100

For Image Classification : HarDNet 78.0 top-1 acc. / 1029.76 Throughput on ImageNet-1K @224x224

For Object Detection : CenterNet-HarDNet 44.3 mAP / 60 FPS on COCO val @512x512

For Semantic Segmentation : FC-HarDNet 75.9% mIoU / 90.24 FPS on Cityscapes test @1024x2048

For Polyp Segmentation : HarDNet-MSEG 90.4% mDice / 119 FPS on Kvasir-SEG @352x352

Main Results

Performance on DFUC2022 Challenge Dataset

We improve HarDNet-MSEG, enhancing its backbone and decoder for DFUC.

Method	DFUC Val. Stage mDice	DFUC Val. Stage mIoU	DFUC Testing Stage mDice	DFUC Testing Stage mIoU
HarDNet-MSEG	65.53	55.22	n/a	n/a
HarDNet-DFUS	70.63	60.49	72.87	62.52

Sample Inference and Visualized Results of FUSeg Challenge Dataset

(Due to the non-disclosure agreement of DFUC2022 dataset, we use another dataset to visualize the results)

HarDNet-DFUS Architecture

Installation & Usage

Environment setting (Prerequisites)

conda create -n dfuc python=3.6
conda activate dfuc
pip install -r requirements.txt

Training

1. Download weights and place in the folder /weights
2. Run:

   python train.py --rect --augmentation --data_path /path/to/training/data
   
   Optional Args:
   --rect         Padding image to square before resize to keep its aspect ratio
   --augmentation Activating data audmentation during training
   --kfold        Specifying the number of K-Fold Cross-Validation
   --k            Training the specific fold of K-Fold Cross-Validation
   --dataratio    Specifying the ratio of data for training
   --seed         Reproducing the result of data spliting in dataloader
   --data_path    Path to training data
   

Testing

Run:

python test.py --rect --weight path/to/weight/or/folder --data_path path/to/testing/data

Optional Args:
--rect         Padding image to square before resize to keep its aspect ratio
--tta          Test time augmentation, 'v/h/vh' for verti/horiz/verti&horiz flip
--weight       It can be a weight or a fold. If it's a folder, the result is the mean of each weight result
--data_path    Path to testing data
--save_path    Path to save prediction mask

Evaluation

Run:

python evaluate.py --image_root path/to/image/folder --gt_root path/to/ground truth/folder

Optional Args:
--image_root    Path to predict result data
--gt_root       Path to ground truth data

Reproduce our best submission in DFUC 2022 Challenge Testing Stage

1. Download the weights for HarDNet-DFUS and place them in the same folder, specifying the folder in --weight when testing. (Please ensure there is no other weight in the folder to obtain the same result.)
2. Run HarDNet-DFUS with 5-fold cross validation and TTA vhflip :

   python test.py --rect --modelname lawinloss4 --weight /path/to/HarDNet-DFUS_weight/folder --data_path /path/to/testing/data --tta vh
   

Acknowledgement

• This research is supported in part by a grant from the Ministry of Science and Technology (MOST) of Taiwan.
  We thank National Center for High-performance Computing (NCHC) for providing computational and storage resources.