Curated Dataset from GRAZPEDWRI-DX
The dataset used in this paper is a curated subset of GRAZPEDWRI‑DX, intended for fine‑grained recognition of wrist pathologies. It includes image data for training, validation, and testing sets.
Dataset description and usage context
Dataset Overview
Data Source
The dataset is carefully selected from GRAZPEDWRI‑DX for research on pediatric wrist pathology recognition. The curated dataset link is: Curated Dataset.
Dataset Structure
The dataset is split into training, validation, and test sets with the following structure:
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
...
Pre‑trained Model
Weights for the refined fine‑grained visual recognition (FGVR) model are available at: Weights.
Evaluation Results
The method performs strongly on the limited test set, with detailed results as follows:
Comparison with 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 Ensemble and FPN Tuning
| Model | Test Set 1 Accuracy (%) | Test Set 2 Accuracy (%) |
|---|---|---|
| PIM | 84.38 | ... |
| PIM + LION | 85.44 | ... |
Pair the dataset with AI analysis and content workflows.
Once the source passes your review, move straight into summarization, transformation, report drafting, or presentation generation with the JuheAI toolchain.