Chest X-ray Disease Classification with DenseNet121

This project trains a deep learning model using transfer learning (DenseNet121) to classify chest X-ray images into 15 categories (14 disease states + "No_Finding") using the NIH Chest X-ray dataset.

---

📁 Dataset

• Source: NIH Chest X-ray dataset (provided on HiPerGator)
• Path: /lustre/fs0/bsc4892/share/ChestXray-NIHCC/images_14_cat
• Classes: 15 total (including "No_Finding")

---

🧠 Model & Approach

• Architecture: DenseNet121 with pretrained ImageNet weights
• Modification: Final classifier layer adjusted to output 15 classes
• Loss Function: CrossEntropyLoss with class balancing via WeightedRandomSampler
• Optimizer: Adam (lr=1e-4)

---

🔁 Training Details

• Image Size: 224 x 224
• Epochs: 5
• Batch Size: 32
• Transforms:
  • Resize
  • ToTensor
  • Normalize (ImageNet stats)

---

📊 Evaluation

• Confusion Matrix: Plotted after model evaluation
• Metrics: Accuracy, Precision, Recall, F1-score via classification_report

---

🧪 Reproducibility

Dependencies

See requirements.txt for all required libraries. Key ones include:

• torch
• torchvision
• matplotlib
• seaborn
• scikit-learn
• tqdm

---

📝 Notes

• The model uses a weighted sampler to handle class imbalance.
• ROC curves can be added by calculating per-class AUC using roc_auc_score from sklearn.

---

📂 Files in this Repository

• chest_xray_densenet.ipynb – Full training and evaluation notebook
• README.md – This file