Chest X-Ray Pneumonia Detection

Classifying chest X-rays as Normal or Pneumonia using DenseNet121 fine-tuned on the Kaggle Chest X-Ray dataset. Added Grad-CAM to visualize which part of the image the model is actually looking at.

---

Results

Metric	Value
Accuracy	95.5%
AUC-ROC	0.9712
Recall	96.2%
F1 Score	0.9650
Precision	95.9%

Detailed breakdown in outputs/reports/evaluation_report.md.

---

Training

Used cosine LR scheduling and early stopping (patience=5). Converges around epoch 20 on GPU, a bit longer on CPU.

---

Confusion Matrix

597 out of 624 test images correct. The main concern in medical screening is false negatives — missing actual pneumonia cases. Got it down to 11.

---

ROC Curve

AUC of 0.97 means the model separates Normal vs Pneumonia well across all possible thresholds, not just the default 0.5 cutoff.

---

Grad-CAM

Grad-CAM shows which regions of the X-ray the model focused on. In the example above, it correctly highlights the consolidation in the right lower lobe rather than something irrelevant like the spine or edges.

---

Prediction Examples

High confidence on clean cases (97–98%). The two misclassifications are borderline cases — a subtle opacity that's easy to miss, and atelectasis that looks similar to early consolidation.

---

How it works

Kaggle dataset  →  resize + augment  →  DenseNet121  →  train  →  evaluate  →  Grad-CAM

More specifically:

• Images resized to 224×224, normalized with ImageNet stats
• Augmentation: horizontal flip, ±10° rotation, brightness/contrast jitter, Gaussian noise
• DenseNet121 pretrained on ImageNet, last layer replaced with a 2-class head
• WeightedRandomSampler to handle the 3:1 class imbalance
• AdamW optimizer, cosine LR schedule, mixed precision on GPU

---

Dataset

Kaggle — Chest X-Ray Images (Pneumonia) — originally from Guangzhou Women and Children's Medical Center.

Split	Normal	Pneumonia	Total
Train	1,208	3,489	4,697
Validation	143	396	539
Test	234	390	624

The original Kaggle val set only has 16 images so data/prepare_data.py moves 10% of train into val.

---

Why DenseNet121

DenseNet121 was used in CheXNet (Rajpurkar et al., 2017) which hit radiologist-level performance on chest X-ray classification. The dense skip connections let gradients flow cleanly to early layers and help the model pick up fine-grained texture differences in lung tissue. ResNet50 is also supported if you want something faster.

---

Setup

pip install -r requirements.txt

Download the dataset:

pip install kaggle
kaggle datasets download -d paultimothymooney/chest-xray-pneumonia
unzip chest-xray-pneumonia.zip -d data/raw/
python data/prepare_data.py

Train:

python main.py --mode train

Evaluate:

python main.py --mode evaluate --checkpoint outputs/models/best_model.pth

Predict on a single image:

python main.py --mode predict --checkpoint outputs/models/best_model.pth --image path/to/xray.jpg

Quick smoke test (no data needed):

python main.py

---

Config

Everything is in configs/config.yaml. Main things you might want to change:

model:
  name: densenet121   # or resnet50

training:
  epochs: 20
  batch_size: 32
  learning_rate: 0.0001

Set num_workers: 0 if you're on Windows and hit DataLoader errors.

---

Saved model

The checkpoint file isn't committed (it's ~30 MB and above what GitHub likes). Either retrain:

python main.py --mode train

or upload outputs/models/best_model.pth somewhere and drop the link here.

---

References

• CheXNet — Rajpurkar et al. (2017) — arXiv:1711.05225
• DenseNet — Huang et al. (2017) — arXiv:1608.06993
• Grad-CAM — Selvaraju et al. (2017) — arXiv:1610.02391