VetFractureAI

Femoral Shaft Fracture Detection & AO-Based Severity Classification

---

Overview

VetFractureAI is a deep learning system based on Faster R-CNN that detects femoral shaft fractures in veterinary X-ray images and classifies their severity using the AO/OTA fracture classification system.

The model performs:

• Fracture localization using bounding boxes
• AO fracture severity classification
• Single-pass inference using Faster R-CNN

Medical Disclaimer: This project is intended for research and educational purposes only and is not a certified medical device. Always consult a licensed veterinarian for clinical diagnosis and treatment decisions.

---

AO Fracture Classification

Class	Label	Description	Surgical Implication
0	No Fracture	Intact femoral shaft	No surgical intervention
1	32-A	Simple fracture (two fragments)	Intramedullary pin or plate fixation
2	32-B	Wedge fracture	More complex fixation required
3	32-C	Complex / comminuted	Advanced fixation techniques

---

Dataset

• Total images: 1,264 veterinary radiographs
• Dogs: 690 images · Cats: 574 images
• Views: Anteroposterior (AP) and lateral
• Annotation tool: CVAT · Format: COCO JSON
• Split: 70% Train / 15% Validation / 15% Test

Species	No Fracture	32-A	32-B	32-C	Total
Dogs	152	200	150	188	690
Cats	150	165	133	126	574
Total	302	365	283	314	1,264

Dataset Access

Publicly available on Hugging Face — no request needed.

huggingface.co/datasets/alvendherfrancisco/VetFractureAI-Dataset

Terms: Academic and non-commercial use only. No redistribution. Citation required in any publication.

---

Model Weights Access

Publicly available on Hugging Face — no request needed.

huggingface.co/alvendherfrancisco/VetFractureAI-Model

Download best_model.pth, place it in the models/ directory, and update the path in the Config class:

Config.MODEL_SAVE_PATH = "/content/drive/MyDrive/.../models/best_model.pth"

Terms: Academic and non-commercial use only. No redistribution. Citation required in any publication.

---

Model Architecture

Component	Details
Architecture	Faster R-CNN
Backbone	ResNet-50 + FPN V2
Pretrained weights	COCO
Input resolution	1024 x 1024 px
Output classes	4 (Background, 32-A, 32-B, 32-C)
Optimizer	SGD
Learning rate	0.005
Momentum / Weight decay	0.9 / 0.0005
Scheduler	CosineAnnealingLR (T_max=35)
Batch size / Epochs	16 / 50
Best checkpoint	Epoch 43
Confidence thresholds	32-A: 0.65, 32-B: 0.45, 32-C: 0.65

---

Preprocessing Pipeline

Step	Operation	Parameters
1	Color conversion	BGR to RGB
2	Resize	LongestMaxSize — max 1024 px
3	Pad	Zero-padding to 1024 x 1024 px
4	CLAHE	clip_limit=3.0, tile_grid=8x8, p=1.0
5	Normalize	mean=0.5, std=0.5

Training augmentations:

Transform	Parameters
HorizontalFlip	p=0.5
ShiftScaleRotate	shift +-6.25%, scale +-10%, rotate +-20°, p=0.5
RandomBrightnessContrast	+-0.3, p=0.3

---

Model Performance

Test Set Results

Metric	Score
mAP@0.5	88.36%
Average IoU	77.69%
Accuracy	77.89%
Macro Precision	77.64%
Macro Sensitivity	77.64%
Macro Specificity	97.83%
Macro F1-Score	77.52%

Per-Class Performance

Class	Precision	Recall	F1
No Fracture	0.918	0.978	0.947
32-A Simple	0.804	0.774	0.789
32-B Wedge	0.565	0.619	0.591
32-C Complex	0.818	0.735	0.774

---

Getting Started

1. Clone the Repository

git clone https://github.com/alvendherfrancisco/VetFractureAI.git
cd VetFractureAI

2. Install Dependencies

pip install -r requirements.txt

3. Run in Google Colab

Click the Open in Colab badge at the top of this page and run all cells in order.

Update Config.ROOT_DIR in the Config class to match your Google Drive folder structure.

4. Notebook Sections

Section	Description
1 — Environment Setup	Install all required packages
2 — Mount Google Drive	Connect to Drive for dataset and checkpoint access
3 — Dataset Distribution	Inspect image counts, species, and AO class breakdown
4 — Complete Training Pipeline	Configure, train, and save the best Faster R-CNN model
5 — TensorBoard Monitoring	Visualize training metrics in real-time
6 — Test Set Evaluation	Evaluate the best checkpoint on the held-out test set
7 — Interactive Gradio Interface	Launch a web UI for real-time inference (optional)

---

Repository Structure

VetFractureAI/
│
├── images/ ← Veterinary radiograph dataset 
│   ├── train/                 
│   ├── val/                  
│   └── test/                  
│
├── annotations/ ← COCO JSON annotation files
│   ├── instances_train.json   
│   ├── instances_val.json
│   └── instances_test.json
│
├── models/ ← Saved model checkpoints                   
│   └── best_model.pth         
│
├── logs/ ← Training logs and evaluation outputs                      
│
├── source_code/ ← Main notebook and training scripts
│   └── fracture_detection_and_ao_classification.ipynb
│
├── requirements.txt
├── LICENSE
└── README.md

---

Contact

Email: alvendherfrancisco01@gmail.com GitHub: github.com/alvendherfrancisco

---

License

Licensed under the Apache 2.0 License. See LICENSE for details.