ADNet

This repository contains PyTorch-based implementation for "Attentional Decoder Networks for Chest X-ray Image Recognition on High-resolution Features." Our ADNet aims to enhance chest X-ray performance via a lightweight attentional decoder network and harmonic magnitude transforms. We provide model, training, and evaluation (supervised, k-NN classifier) code as well as pre-trained weight in the hope that researchers and practitioners widely use our works effortlessly.

Method Overview of Attentionl Decoder Network

Tutorial

1️⃣ [Inference Only]

1. Copy mecla/model/convnext.py into your project root folder.

2. Install timm==0.6.2 and run the following code snippets. We provide the following models with checkpoints.

   • Model Name patterns: [resnet50|convnext_tiny]_ADNet_[Dataset Name]
   • Dataset Name: IN1K, IN1K_NIH, IN1K_MIMIC, IN1K_CheXpert, ALL, ALL_NIH, ALL_MIMIC, ALL_CheXpert.

   !wget -nc https://raw.githubusercontent.com/Lab-LVM/ADNet/main/mecla/model/convnext.py
   !wget -nc https://raw.githubusercontent.com/Lab-LVM/ADNet/main/misc/sample.png
   import convnext
   from timm import create_model
   from PIL import Image
   from torchvision.transforms import Compose, Resize, CenterCrop, Normalize, ToTensor, Grayscale
   
   img = Image.open('sample.png')
   x = Compose([
       Grayscale(3), Resize(472, 3), CenterCrop(448), ToTensor(), 
       Normalize((0.485, 0.456, 0.406), (0.229, 0.224, 0.225))
   ])(img)
   model = create_model('convnext_tiny_ADNet_ALL_NIH', num_classes=14, pretrained=True)
   model.eval()
   y = model(x.unsqueeze(0)).argmax()
   print(y)
   # output: tensor(1) (cardiomegaly)
   

2️⃣ [Train]

1. Clone this repository and install the dependency.

   git clone https://github.com/Lab-LVM/ADNet
   pip install -r requirements.txt
   

2. Download chest X-ray datasets and change data_dir in config/train.json .

   "data_dir": {
     "chexpert": "/path/to/CheXpert-v1.0-small/",
     "nihchest": "/path/to/nih/",
     "mimic": "/path/to/mimic-cxr/",
   }
   

   You can download each dataset using the provided URLs, and the dataset directory looks like the below:

   Tip.

   • We resize images to 384x384 resolution using misc/resize_chexpert.py.
   • We split MIMIC-CXR using CheXclusion protocols. You can download the test patient ID lists from here. Specifically, we upload our dataset preparation guideline for mimix-cxr to here.

   # 1. CheXpert: https://stanfordmlgroup.github.io/competitions/chexpert/
   CheXpert-v1.0-small
   ├── train
   ├── train.csv
   ├── valid
   └── valid.csv
   
   # 2. MIMIC-CXR: https://physionet.org/content/mimic-cxr/2.0.0/
   mimic-cxr
   ├── files
   ├── img_384
   ├── test_x.json
   ├── test_y.json
   ├── train_x.json
   └── train_y.json
   
   # 3. https://www.kaggle.com/datasets/nih-chest-xrays/data
   nih
   ├── Data_Entry_2017.csv
   ├── img_384
   ├── test_list.txt
   └── train_val_list.txt
   

3. Run the following command.

   • Download a pre-trained checkpoint and modify the checkpoint path of model_weight in config/train.json. You can choose a checkpoint depending on pre-trained dataset types: ImageNet or ALL (NIH+MIMIC+CheXpert).

     No	Backbone	IN1K	ALL
     1	ResNet50	ckpt	ckpt log
     2	ConvNeXt-T	ckpt	ckpt log

   • See more training scripts in script/supervised/.

   • See available training settings in config/train.json.

   • Add --use-wandb for watching progress in wandb.

   # pattern: python3 train.py -c [GPU Device] -s [Setting Name] -m [Model Name]
   python3 train.py -s nihchest -c 1 -m convnext_tiny_ADNet_IN1K
   

Experiment Result

We provide experiment results with pre-trained checkpoints.

Pre-train Data	Fine-tune Data	Backbone	Image	AUROC	Download
IN1K	NIH	ResNet50	448	83.5
IN1K	NIH	ConvNeXt-T	448	83.8	ckpt log
IN1K	MIMIC	ResNet50	256	84.5
IN1K	MIMIC	ConvNeXt-T	416	85.3	ckpt log
IN1K	CheXpert	ResNet50	256	90.6
ALL	NIH	ResNet50	448	84.1	ckpt log
ALL	NIH	ConvNeXt-T	448	84.4	ckpt log
ALL	MIMIC	ResNet50	448	85.2	ckpt log
ALL	MIMIC	ConvNeXt-T	448	85.6	ckpt log

Note

• In the pre-train dataset column, ALL means a dataset combined by NIH, MIMIC, and CheXpert using misc/combine_datasets.py.
• We run our experiments on Ubuntu 18.04 LTS, CUDA 11.03, and a single NVIDIA RTX 3090 24GB GPU.

Acknowledgment

Our code is based on the TIMM library.