PneumoScan: Pneumonia Detection from Chest X-Rays (v1 – Baseline)

🩻 Built by an Ex-Surgeon turned ML Engineer
A simple yet powerful deep learning model for detecting pneumonia from chest X-ray images.

Left: Normal lungs | Right: Pneumonia with consolidation (source: educational medical video)

Project Overview

This is the first version (baseline) of PneumoScan – a binary classification model that distinguishes NORMAL from PNEUMONIA in chest X-rays.

• Dataset: Kaggle Chest X-Ray Images (Pneumonia) – ~5,863 images
• Model: Pretrained EfficientNet-B0 (transfer learning)
• Achieved Test Accuracy: 94.71% on hold-out test set
• Precision/Recall/F1 strong for pneumonia class (critical for clinical use)

As a former surgeon (with experience in thoracic cases), 10+ years mobile developer (Android/iOS), and self-taught ML engineer (TensorFlow Certified), I built this to create clinically relevant AI tools.

Key Results

• Test Accuracy: 0.9471 (94.71%)
• Classification Report:

Demo

Try it live!

Here is the confusion matrix from the test set:

Similar pattern to our model: High recall for Pneumonia, low false negatives

More sample X-ray comparisons:

Clear visual difference: Pneumonia shows opacities/consolidation

Tech Stack

• Framework: PyTorch 2.0+
• Model: torchvision.models.efficientnet_b0 (pretrained on ImageNet)
• Data: torchvision.datasets.ImageFolder + transforms
• Optimizer: AdamW (lr=3e-4, weight decay=1e-4)
• Scheduler: CosineAnnealingLR
• Evaluation: sklearn (classification_report, confusion_matrix), seaborn/matplotlib
• Training: 12 epochs on GPU (~15-25 min)

Dataset

• Source: Kaggle – Chest X-Ray Images (Pneumonia)
• Structure: train/val/test folders with NORMAL and PNEUMONIA subfolders
• Classes: ['NORMAL', 'PNEUMONIA']