A Quality Control Model for Wrist X-ray Images

Visualization of model prediction results

Visualization of model evaluation results

Introduction

The project pipeline includes one classification model and two landmarks detection models. At the same time, we checked the relevant information in the image file and left and right markers on images.
classification referring to this paper: MobileNetV2: Inverted Residuals and Linear Bottlenecks
landmarks detection referring to this paper: You Only Learn Once: Universal Anatomical Landmark Detection

Requirements:

python==3.9.12
albumentations==1.3.1
matplotlib==3.7.1
numpy==1.24.3
pydicom==2.4.0
PyYAML==6.0
PyYAML==6.0
scikit_learn==1.2.2
skimage==0.0
torch==2.0.1
torchsummary==1.5.1
torchvision==0.15.2

Code Structure

.
├── augmentation
│   ├── __init__.py
│   └── medical_augment.py
├── checkpoints
│   ├── model_best_wrist-classify-artifact.pth.tar
│   ├── model_best_wrist-classify-overlap.pth.tar
│   ├── model_best_wrist-classify-position.pth.tar
│   ├── model_best_wrist-landmarks-AP.pth.tar
│   └── model_best_wrist-landmarks-LAT.pth.tar
├── configs
│   ├── config_classify_artifact.yaml
│   ├── config_classify_overlap.yaml
│   ├── config_classify_position.yaml
│   ├── config_inference.yaml
│   ├── config_landmarks_AP.yaml
│   └── config_landmarks_LAT.yaml
├── data
│   ├── cp_mv.py
│   ├── gen_list_meta_position.py		# generate .txt about the name of files and label/coordinates
│   ├── gen_list_meta.py
│   ├── wrist		        # folder to save .png & .json
│   ├── wrist_AP		# folder to save .png & .json
│   └── wrist_LAT		# folder to save .png & .json
├── dataset
│   ├── __init__.py
│   ├── xray_classify_dataset.py
│   └── xray_landmark_dataset.py
├── depoly
│   ├── infer_with_tensorRT.py
│   ├── onnx_model_202307131414.onnx
│   ├── torch2onnx.py
│   └── trt_model.trt
├── eval		# Scoring by information in checking, mark of position and coordinates
│   ├── char_recognize.py
│   ├── chkmsg_judge.py
│   ├── __init__.py
│   └── point_judge.py
├── example_eval.png
├── example_infer.png
├── experiments
│   ├── wrist-classify-artifact
│   │   ├── indicators_of_valid.txt
│   │   └── log.txt
│   ├── wrist-classify-overlap
│   │   ├── indicators_of_valid.txt
│   │   └── log.txt
│   ├── wrist-classify-position
│   │   ├── indicators_of_valid.txt
│   │   └── log.txt
│   ├── wrist-landmarks-AP
│   │   ├── indicators_of_valid.txt
│   │   ├── log0808.txt
│   │   ├── pred_filenames.txt
│   │   └── pred_landmarks.txt
│   └── wrist-landmarks-LAT
│       ├── indicators_of_valid.txt
│       ├── pred_filenames.txt
│       └── pred_landmarks.txt
├── files_from_hos              	 	# folder to save .dcm
├── inference.py
├── inference_result
│   ├── DX145789.png
│   ├── DX148722.png
│   └── inference.json
├── inference_task              		# folder to save .dcm (files needed to be inferred)
│   ├── DX145789.dcm
│   └── DX148722.dcm
├── losses
│   ├── __init__.py
│   └── loss.py
├── models
│   ├── densenet.py
│   ├── GLNet.py
│   ├── globalNet.py
│   ├── __init__.py
│   ├── mobilenet.py
│   ├── resnet.py
│   └── unet_dw.py
├── README.md
├── read_result.py
├── requirements.txt
├── tools
│   ├── char-L0.png
│   ├── char-L1.png
│   ├── char-R0.png
│   ├── char-R1.png
│   ├── manual_evaluation_AP.xlsx
│   ├── classify_AP_LAT.py
│   ├── compare_manual_model_AP.py
│   ├── compare_manual_model_LAT.py
│   ├── dcm2png.py
│   ├── find_range.py
│   └── form_matched_image.py
├── train_classify.py
├── train_landmark.py
├── train.sh
├── utils
│   ├── heatmap.py
│   ├── __init__.py
│   ├── logger.py
│   ├── misc.py
│   └── progress_bar.py
└── wrist_data_dcm		# after executing 'classify_AP_LAT.py', dcm files will be moved here from 'files_from_hos'
    ├── wrist_AP
    └── wrist_LAT

Quickly Start

1. Place DICOM files in directory files_from_hos
2. executing classify_AP_LAT.py, DICOM files will be moved to wrist_data_dcm (please modify the code as needed)
3. executing tools/dcm2png.py, then executing data/gen_list_meta.py after completing annotation (I used labelme)
4. modify the config file experiments/config_xxx.yaml
5. Training Model: train_classify.py train_landmark.py
6. Scoring in quality control: inference.py

welcome to open an issue to communicate with me.