MVP-Net: Multi-view FPN with Position-aware Attention for Deep Universal Lesion Detection

This is an implementation of MICCAI 2019 paper MVP-Net: Multi-view FPN with Position-aware Attention for Deep Universal Lesion Detection.

Installation

This code is based on Detectron.pytorch. Please see it for installation.

Environment

• Python (Tested on 3.6)
• PyTorch (Tested on 0.4.1.post2)

Data preparation

Download DeepLesion dataset here.

We provide coco-style json annotation files converted from DeepLesion. Unzip Images_png.zip and make sure to put files as following sturcture:

data
  ├──DeepLesion
        ├── annotations
        │   ├── deeplesion_train.json
        │   ├── deeplesion_test.json
        │   ├── deeplesion_val.json
        └── Images_png
              └── Images_png
               │    ├── 000001_01_01
               │    ├── 000001_03_01
               │    ├── ...

Training

To train MVP-Net with 9 slices model, run:

bash multi_windows_9_slices.sh train

We also provide our re-implementation of 3DCE, see 3DCE_*.sh for training and testing.

Testing

After training, put the model path into .sh file, after '--load_ckpt', and run:

bash multi_windows_9_slices.sh test

Results on DeepLesion dataset

FPs per image	0.5	1	2	3	4
ULDOR	52.86	64.80	74.84	-	84.38
3DCE, 3 slices	55.70	67.26	75.37	-	82.21
3DCE, 9 slices	59.32	70.68	79.09	-	84.34
3DCE, 27 slices	62.48	73.37	80.70	-	85.65
FPN+3DCE, 3 slices*	58.06	68.85	77.48	81.03	83.27
FPN+3DCE, 9 slices*	64.25	74.41	81.90	85.02	87.21
FPN+3DCE, 27 slices*	67.32	76.34	82.90	85.67	87.60
Ours, 3 slices	70.01	78.77	84.71	87.58	89.03
Ours, 9 slices	73.83	81.82	87.60	89.57	91.30
Imp over 3DCE, 27slices	11.35	8.45	6.90	-	5.65

* indicates our re-implementation of 3DCE with FPN as backbone.

Contact

If you have questions or suggestions, please open an issue here or send an email to lizihao2018@ia.ac.cn.