Deep Learning Chest X-Ray Classifier

Comparative Study: Custom CNN vs. Transfer Learning for Pneumonia Detection

A deep learning project focused on binary classification of chest X-ray images (Normal vs. Pneumonia) using two different approaches:

• A custom Convolutional Neural Network (CNN) built from scratch
• A Transfer Learning pipeline using EfficientNetB0

The project emphasizes not only model performance, but also engineering realism, including:

• prevention of data leakage
• handling of hardware limitations
• medically appropriate preprocessing choices
• practical deployment considerations for constrained environments

---

Overview

Most beginner medical imaging repositories unintentionally introduce methodological flaws such as:

• validating directly on the test set
• unrealistic augmentation policies
• ignoring hardware execution limitations
• presenting transfer learning as universally plug-and-play

This repository intentionally avoids those mistakes.

The goal was to build a more technically honest comparison between:

1. a lightweight custom CNN that can reliably train on CPU hardware
2. a modern transfer learning architecture that exposes real-world execution constraints

---

Key Features

Data Integrity First

A strict dataset separation strategy was enforced:

• 80% Training
• 20% Validation
• Completely isolated Test Set

This prevents data leakage, which is extremely common in introductory ML projects where validation performance becomes artificially inflated.

---

Medical Context Awareness

Medical image augmentation was designed carefully.

Disabled Augmentations

• Horizontal flipping was intentionally disabled to preserve anatomical correctness.
• Random transformations were kept conservative to avoid unrealistic pathology distortion.

Why this matters:

• Flipping chest X-rays can create anatomically impossible conditions such as situs inversus
• Aggressive augmentations can introduce medically invalid features

---

Project Structure

├── Solution_1.py   # EfficientNetB0 Transfer Learning
├── Solution_2.py   # Custom CNN Baseline
├── dataset/
├── models/
├── results/
└── README.md

---

Solution 2 — Production Baseline (Custom CNN)

Why This Became the Primary Model

The custom CNN became the repository’s primary working solution because it trained reliably under limited hardware conditions.

Unlike the transfer learning implementation, it:

• converged consistently
• generalized properly
• executed without numerical instability on CPU-only systems

---

Architecture

Convolution Blocks

Progressive feature extraction:

32 → 64 → 128 → 128

Pooling Strategy

Instead of large dense layers, the network uses:

GlobalAveragePooling2D()

Benefits:

• lower parameter count
• reduced overfitting
• improved generalization

---

Regularization

Heavy regularization was necessary due to:

• relatively small medical dataset
• high overfitting risk

Implemented techniques:

• Dropout(0.5)
• Data augmentation
• Early stopping

---

Outcome

The custom CNN:

• trained successfully
• achieved stable validation performance
• demonstrated proper generalization behavior
• validated the integrity of the preprocessing pipeline

This model serves as the repository’s fully functional reference implementation.

---

Solution 1 — Transfer Learning with EfficientNetB0

Objective

The goal was to leverage:

• ImageNet pre-trained weights
• EfficientNetB0 feature extraction capabilities
• fine-tuning for chest X-ray classification

This implementation uses:

• Keras Functional API
• frozen EfficientNet backbone
• custom classification head

---

Important Engineering Detail

To avoid catastrophic forgetting during fine-tuning:

training=False

was explicitly enforced for Batch Normalization layers inside EfficientNet.

This is considered best practice when adapting pre-trained CNN architectures to smaller domain-specific datasets.

---

The Real Problem: Hardware Constraints

This repository intentionally documents a major practical issue that many tutorials ignore.

Environment

• Native Windows environment
• CPU-only execution
• TensorFlow versions > 2.10
• No CUDA acceleration

---

Observed Failure

Despite correct implementation:

• training appeared to execute normally
• loss calculations became numerically unstable internally
• gradients effectively vanished
• validation accuracy became permanently stuck at:

50%

This occurred consistently during EfficientNet execution.

---

Root Cause

The issue was not architectural logic.

The bottleneck was:

• CPU-only execution
• TensorFlow backend instability on native Windows
• computational overhead from EfficientNet graph operations

In practice, complex transfer learning pipelines are heavily dependent on:

• CUDA-enabled GPUs
• Linux-based deep learning environments
• optimized tensor computation libraries

---

Key Takeaway

This repository demonstrates an important engineering reality:

A theoretically correct deep learning architecture can still fail due to infrastructure limitations.

Transfer learning is not automatically superior if:

• the environment cannot execute it reliably
• numerical stability collapses during training
• hardware constraints dominate experimentation

Under constrained hardware:

• simpler architectures often outperform complex models in practice
• reproducibility matters more than theoretical sophistication

---

Technologies Used

• Python
• TensorFlow / Keras
• NumPy
• Matplotlib
• Scikit-learn

---

Future Improvements

Potential extensions include:

• Linux + CUDA migration
• Mixed precision training
• Grad-CAM visualization
• Multi-class disease classification
• Hyperparameter optimization
• Cross-validation benchmarking
• Model deployment with Flask/FastAPI

---

How to Run

Install Dependencies

pip install tensorflow numpy matplotlib scikit-learn

---

Run Custom CNN

python Solution_2.py

---

Run Transfer Learning Experiment

python Solution_1.py

---

Disclaimer

This project is intended for:

• educational purposes
• engineering experimentation
• comparative deep learning analysis

It is not a clinical diagnostic system and should not be used for medical decision-making.