HarmeetKalha/pneumonia-detector

Chest X-Ray Pneumonia Detection with Grad-CAM

Fine-tuned EfficientNet-B0 to detect pneumonia from chest X-rays, with Grad-CAM explainability implemented from scratch to visualise model attention regions.

---

Results

Metric	Value
Test Accuracy	93%
ROC-AUC	0.9798
Pneumonia F1	0.94
Pneumonia Recall	0.97

Evaluated on 624 test images (234 NORMAL, 390 PNEUMONIA).

---

Grad-CAM Visualisations

Grad-CAM (Gradient-weighted Class Activation Mapping) highlights which regions of the X-ray the model focused on when making its prediction. Implemented from scratch without external libraries.

Finding: The model primarily attends to lung tissue for pneumonia classification. Some attention to peripheral regions (arms) suggests potential shortcut learning — a deployment robustness concern relevant to clinical settings.

---

Training Curves

---

ROC Curve

---

Confusion Matrix

---

Dataset

Chest X-Ray Images (Pneumonia) by Paul Mooney on Kaggle.

• 5,216 training images
• 624 test images
• Classes: NORMAL, PNEUMONIA
• Class imbalance: ~3:1 (PNEUMONIA:NORMAL)

---

Key Design Decisions

Transfer Learning EfficientNet-B0 pretrained on ImageNet. All layers fine-tuned at a low learning rate (1e-4) to preserve pretrained features while adapting to chest X-ray domain.

Handling Class Imbalance Two complementary approaches:

• WeightedRandomSampler — oversamples minority class during training
• Weighted CrossEntropyLoss — penalises misclassification of minority class more heavily

Learning Rate Scheduling ReduceLROnPlateau reduces LR by 10x if validation loss plateaus for 3 epochs — automatically adapts without manual tuning.

Grad-CAM Implementation Forward hook captures feature maps at EfficientNet's last conv block. Backward hook captures gradients. Weights each feature map by its mean gradient, takes weighted sum, applies ReLU, upsamples to image size. No external explainability libraries used.

---

Project Structure

chest_xray_project/
├── model.py          — EfficientNet-B0 with custom classification head
├── train.py          — training loop, weighted sampler, LR scheduler
├── evaluate.py       — metrics, confusion matrix, ROC curve
├── gradcam.py        — Grad-CAM implementation from scratch
├── visualise.py      — heatmap generation and overlay
├── outputs/          — training curves, confusion matrix, ROC curve
└── gradcam_outputs/  — heatmap visualisations

---

Setup

git clone https://github.com/HarmeetKalha/chest-xray-pneumonia
cd chest-xray-pneumonia
pip install torch torchvision matplotlib seaborn scikit-learn opencv-python

Download the dataset from Kaggle and place it at:

data/chest_xray/train/
data/chest_xray/val/
data/chest_xray/test/

Download model weights from Hugging Face and place in checkpoints/best_model.pth.

---

Training

python train.py

Evaluation

python evaluate.py

Grad-CAM Visualisation

python visualise.py

---

Author

Harmeet Singh Kalha

• GitHub: @HarmeetKalha
• LinkedIn: harmeetsinghkalha