Learning from the Few: Fine-grained Approach to Pediatric Wrist Pathology Recognition on a Limited Dataset

Paper URL: https://www.sciencedirect.com/science/article/pii/S0010482524011296

We approach wrist pathology recognition as a fine-grained recognition problem. We refine our fine-grained architecture through ablation analysis and the integration of LION optimization. Leveraging Grad-CAM, an explainable AI technique, we highlight these regions. Despite using limited data, reflective of real-world medical study constraints, our method consistently outperforms state-of-the-art image recognition models on both augmented and original (challenging) test sets. Our proposed refined architecture achieves an increase in accuracy of 1.06% and 1.25% compared to the baseline method, resulting in accuracies of 86% and 84%, respectively. Moreover, our approach demonstrates the highest fracture sensitivity of 97%, highlighting its potential to enhance wrist pathology recognition.

Evaluation Against Other Deep Neural Networks

| Model | Test Accuracy (%) | |----------------|-------------------| | EfficientNetV2 | 53.59 | | NFNet | 65.40 | | VGG16 | 65.82 | | ViT | 70.25 | | DeiT3 | 70.89 | | RegNet | 72.36 | | DenseNet201 | 73.42 | | MobileNetV2 | 76.37 | | CMAL | 76.58 | | RexNet100 | 77.43 | | ResNet101 | 77.43 | | IELT | 78.10 | | DenseNet121 | 78.21 | | ResNest101e | 78.27 | | InceptionV4 | 78.69 | | MetaFormer | 78.90 | | ResNet50 | 79.11 | | InceptionV3 | 79.54 | | EfficientNet_b0 | 79.96 | | YOLOv8x | 80.50 | | HERBS | 82.70 | | Our Approach (PIM for FGVR) | 84.38 |

LION Integration and FPN Adjustment

| Model | Test Set 1 Accuracy (%) | Test Set 2 Accuracy (%) | |----------------|-------------------------|-------------------------| | PIM | 84.38 | 82.50 | | PIM + LION | 85.44 | 83.75 |

1. Environment setting

1.0. Package

• Install Requirements pip install -r requirements.txt

1.1. Dataset

In this paper, we use a dataset curated from GRAZPEDWRI-DX. The curated dataset can be found at: * Curated Dataset

1.2. Our Pre-trained Model

The weights of the refined FGVR model can be found at: * Weights

1.3. OS

• [x] Windows10
• [x] Ubuntu20.04
• [x] macOS (CPU only)

2. Train

• [x] Single GPU Training
• [x] DataParallel (single machine multi-gpus)
• [ ] DistributedDataParallel

(more information: https://pytorch.org/tutorials/intermediate/ddp_tutorial.html)

2.1. Data

train, val, and test data structures:
├── train/ │ ├── 0/ │ | ├── 0/0133_0306769778_07_WRI-R2_M015-1.png │ | ├── 0133_0306769778_07_WRI-R2_M015-3.png │ | └── .... │ ├── 1/ │ | ├── 0025_0483842914_01_WRI-L2_F000.png │ | ├── 0053_1119833109_03_WRI-R1_F005.png │ | └── .... │ └── .... └── val/ │ ├── 0/ │ | ├── 0133_0306769778_07_WRI-R2_M015-0.png │ | ├── 0133_0306769778_07_WRI-R2_M015-2.png │ | └── .... │ ├── 1/ │ | ├── 0042_0827512771_04_WRI-R2_M015.png │ | ├── 0071_0680563744_02_WRI-R1_F009.png │ | └── .... │ └── .... └── test/ │ ├── 0/ │ | ├── 0772_0547017117_03_WRI-R1_M017-0.png │ | ├── 0772_0547017117_03_WRI-R1_M017-1.png │ | └── .... │ ├── 1/ │ | ├── 0069_0502540283_01_WRI-L1_M013.png │ | ├── 0078_1212376595_01_WRI-L1_M011.png │ | └── .... │ └── .... └── test2/ │ ├── 0/ │ | ├── 0772_0547017117_03_WRI-R1_M017.png │ | ├── 0834_0240036198_01_WRI-R1_M014.png │ | └── .... │ ├── 1/ │ | ├── 0069_0502540283_01_WRI-L1_M013.png │ | ├── 0115_0432451427_01_WRI-L2_M004.png │ | └── .... │ └── ....

2.2. Configuration

All the important configuration files are located at config folder. You can directly modify yaml file if needed.

2.3. Run

python main.py --c ./configs/wrist.yaml model will save in ./records/{projectname}/{expname}/backup/

2.5. Multi-GPUs

comment out main.py line 66 model = torch.nn.DataParallel(model, device_ids=None)

3. Evaluation

If you want to evaluate our pre-trained model or your custom-trained model (specify the path to your model in the config file): python main.py --c ./configs/eval.yaml Results will be saved in ./records/{projectname}/{expname}/eval_results.txt

4. Heatmap Generation

python heat.py --c ./configs/wrist.yaml --img data/test/1/0069_0502540283_01_WRI-L1_M013.png --save_img ./0069_0502540283_01_WRI-L1_M013.png

5. Inference

If you want to reason your picture and get the confusion matrix, please provide configs/eval.yaml python infer.py --c ./configs/eval.yaml Results will be saved in ./records/{projectname}/{expname}/infer_results.txt

Citation

If you find our paper useful in your research, please consider citing: @article{AHMED2024Learning, title = {Learning from the few: Fine-grained approach to pediatric wrist pathology recognition on a limited dataset}, author = {Ammar Ahmed and Ali Shariq Imran and Zenun Kastrati and Sher Muhammad Daudpota and Mohib Ullah and Waheed Noor}, journal = {Computers in Biology and Medicine}, volume = {181}, pages = {109044}, year = {2024}, issn = {0010-4825}, doi = {https://doi.org/10.1016/j.compbiomed.2024.109044}, url = {https://www.sciencedirect.com/science/article/pii/S0010482524011296}, }

Acknowledgment

• Thanks to PIM for their base Pytorch implementation of Plug-in Module.

• This work was supported in part by the Department of Computer Science (IDI), Faculty of Information Technology and Electrical Engineering, Norwegian University of Science and Technology (NTNU), Gjøvik, Norway; and in part by the Curricula Development and Capacity Building in Applied Computer Science for Pakistani Higher Education Institutions (CONNECT), Project number: NORPART-2021/10502, funded by DIKU.