Fracture Morphology Classification: Local Multiclass Modeling for Multilabel Complexity

Overview

In this work, we analyse the classification of fracture morphology of paediatric wrist fractures within the AO/OTA system. Therefore, we propose a method to extract fracture morphology by assigning global AO codes to corresponding fracture bounding boxes. This enables a local multi-class classification of single fracture bounding boxes.

Environment

Our code is implemented in PyTorch using the PyTorch Lightning framework. Please use the provided requirements file (requirements.txt) file to create the environment.

python3 -m venv venv
source venv/bin/activate
pip install -r requirements.txt

Data

Please download the dataset using the provided link in the original paper and preprocess it with their provided notebooks to obtain the 8-bit images and store them in the data/preprocessed_images folder. Place the metadata file in the dataset.csv file in the data folder. Place the given ground truth bounding boxes in GRAZPEDWRI-DX/folder_structure/yolov5/labels in the data/gt_bboxes folder.

For the generation of the AO code mapping to their bounding boxes, bone segmentation masks are needed. Please download them using the provided link in the original paper and store them in data/raw_segmentations_all.h5. The Yolo generated bounding boxes are already given in data/yolo_bbox. Please unzip pred0.zip, pred1.zip and pred2.zip anf store the .txt-files directly in data/yolo_bbox.

For running the models, the extraction of the assignment of the global AO codes to the corresponding fracture bounding boxes must be performed by running

python preprocessing/add_ao_codes_to_labels.py

Therefore, the paths at the beginning of the script must be adapted. If the ground truth labels should be generated, set the output_path to data/gt_bboxes_labels and the bool yolo=False. For the yolo bounding boxes set the output_path to data/yolo_bbox_labels and the bool yolo=True.

Training

To train the different the different models, you can run the different approaches presented in the paper in experiments just by if necessary adapting the paths on the top of the scripts and then just run the files by :

python filename.py

Following models are given:

• binary_fracture_classification.py : Binary fracture classification trained on the Yolo bounding boxes
• multilabel_on_full_image.py: Multi-label classification on the full-image on the ground truth AO codes
• multiclass_on_gt_bbox.py: Multi-class classification on the ground truth bounding boxes
• yolo_bbox_classification.py: Multi-class classification on the Yolo bounding boxes
• yolo_bbox_fp_reduction_classification.py: Multi-class classification on the Yolo bounding boxes with the additional Healthy class generated by false-positive reduction.

Evaluation

The evaluation of the models is already done in the training scripts, so just check out the generated results directory. In plotting_results there are two notebooks given for generating a bar plot of the different models showing their F1-Scores and one plot showing the influence of the confidence score for the Yolo approaches.