MURA Model Classifier

Automated fracture detection in musculoskeletal X-ray images from the MURA dataset using deep learning and computer vision techniques.

This repository provides:

• a ready-to-use PyTorch inference package,
• pretrained per-anatomical-area models,
• and a full analytical report describing the methodology and experiments.

The project was developed as an academic research work and is intended for educational and research purposes only.

---

Project Overview

Musculoskeletal Radiographs (MURA) present several challenges for automated analysis:

• high intra-class variability,
• subtle fracture patterns,
• strong anatomical differences between body parts.

To address this, the project follows a per-area modeling strategy: each anatomical region is trained with its own dedicated classifier.

Supported anatomical areas

• XR_ELBOW
• XR_FINGER
• XR_FOREARM
• XR_HAND
• XR_HUMERUS
• XR_SHOULDER
• XR_WRIST

Each area has an independently trained model and checkpoint.

---

Methodology Summary

The full methodology is described in detail in the analytical note, but the main steps are:

1. Dataset preparation

   • MURA dataset with image-level and study-level labels
   • Explicit separation by anatomical area
   • Study-aware aggregation for evaluation

2. Baseline architecture

   • DenseNet-121 pretrained on ImageNet
   • Binary classification head (fracture / no fracture)

3. Training strategy

   • Area-specific training
   • Class imbalance handling
   • Two-phase fine-tuning strategy
   • Best model selection based on validation metrics

4. Hyperparameter optimization

   • Automated tuning using Optuna
   • Separate studies per anatomical area

5. Evaluation

   • Image-level metrics
   • Study-level aggregation and AUC evaluation

6. Interpretability

   • Grad-CAM based visualization
   • Qualitative localization analysis of fracture regions

7. Advanced experiments (research stage)

   • CAM-based regularization
   • Hybrid architectures (DenseNet + Transformer backbones)

---

Analytical Note

A detailed analytical report describing the dataset analysis, experimental setup, training strategies, evaluation protocols, and results is available here:

Analytical Note (PDF)

The document includes:

• MURA dataset analysis,
• per-area classification rationale,
• DenseNet baseline and improvements,
• Optuna hyperparameter optimization,
• CAM / Grad-CAM interpretability analysis,
• two-phase training strategy,
• hybrid DenseNet + Transformer experiments,
• quantitative and qualitative results.

---

Installation

Install directly from GitHub:

pip install git+https://github.com/mariamvol/mura-model-classifier.git

---

Requirements:

• Python ≥ 3.9
• PyTorch
• torchvision
• NumPy, pandas, scikit-learn

---

Inference Usage

Single image inference

mura-infer \
  --image path/to/image.png \
  --area XR_WRIST

On first run, the corresponding pretrained weights will be automatically downloaded from the GitHub release and cached locally.

---

Using custom checkpoint

mura-infer \
  --image path/to/image.png \
  --area XR_WRIST \
  --ckpt /path/to/XR_WRIST_FINAL_best.pt

---

Output

The command prints:

• predicted fracture probability (sigmoid output),
• binary prediction using a configurable threshold (default: 0.5).

Example output:

area=XR_WRIST
prob_fracture=0.8731
pred=1

---

Model Details

• Architecture: DenseNet-121
• Input resolution: 224 × 224
• Normalization: ImageNet statistics
• Output: single logit → sigmoid probability

Each anatomical area uses its own independently trained model.

---

Disclaimer

This project is provided for research and educational purposes only.

It is not a medical device and must not be used for clinical decision-making or diagnostic purposes.

---

License

This project is released under the MIT License.