Skip to main content

​
Neural Network Type

Convolutional Neural Network (CNN) We chose a custom CNN architecture with three convolutional blocks designed specifically for medical image classification.

​
Architecture Justification

​
Why CNN vs Other Neural Network Types?

Problems with MLP:
  • Would require flattening the image (224×224×3 = 150,528 input parameters)
  • Loss of spatial information
  • Prohibitive number of parameters
  • Cannot capture local relationships between pixels
Advantages of CNN:
  • Maintains 2D spatial structure
  • Shared parameters in filters
  • Detects local features (edges, textures)
  • Translation invariance
Problems with RNN/LSTM:
  • Designed for sequential data (text, time series)
  • Don’t leverage spatial structure of images
  • Slower to train
  • Unnecessarily complex for this problem
Advantages of CNN:
  • Optimized for data with spatial structure
  • Efficient parallel operations
  • Proven architecture for medical images

​
CNN Architecture

​
High-Level Design

INPUT (224×224×3) → CONV → POOL → CONV → POOL → CONV → POOL → FC → OUTPUT (2)

​
Detailed Architecture Diagram

┌─────────────────────┐
│  Input: 224×224×3   │
└──────────┬──────────┘
           │
┌──────────▼──────────┐
│  Conv2D (32, 3×3)   │
│  ReLU               │
└──────────┬──────────┘
           │
┌──────────▼──────────┐
│  MaxPool2D (2×2)    │
└──────────┬──────────┘
           │
┌──────────▼──────────┐
│  Conv2D (64, 3×3)   │
│  ReLU               │
└──────────┬──────────┘
           │
┌──────────▼──────────┐
│  MaxPool2D (2×2)    │
└──────────┬──────────┘
           │
┌──────────▼──────────┐
│  Conv2D (128, 3×3)  │
│  ReLU               │
└──────────┬──────────┘
           │
┌──────────▼──────────┐
│  MaxPool2D (2×2)    │
└──────────┬──────────┘
           │
┌──────────▼──────────┐
│  Flatten            │
└──────────┬──────────┘
           │
┌──────────▼──────────┐
│  Dense (128)        │
│  ReLU               │
└──────────┬──────────┘
           │
┌──────────▼──────────┐
│  Dropout (0.5)      │
└──────────┬──────────┘
           │
┌──────────▼──────────┐
│  Dense (2)          │
│  Softmax            │
└──────────┬──────────┘
           │
           ▼
    [Normal, Pneumonia]

​
Layer-by-Layer Breakdown

​
Input Layer

  • Dimension: 224 × 224 × 3 (RGB)
  • Preprocessing: Normalization [0, 1], resizing

​
Convolutional Block 1

Conv2D Layer
  • 32 filters, 3×3 kernel, ReLU activation
  • Why 32 filters? Balance between capacity and efficiency for initial feature detection
  • Why 3×3 kernel? Captures basic local patterns (edges, lines)
MaxPooling2D
  • 2×2 pool size
  • Reduces dimensions to 112×112
  • Provides invariance to small translations
  • Reduces computational load

​
Convolutional Block 2

Conv2D Layer
  • 64 filters, 3×3 kernel, ReLU activation
  • Deeper filters capture more complex patterns
  • Detects textures and opacity patterns
MaxPooling2D
  • 2×2 pool size
  • Reduces dimensions to 56×56

​
Convolutional Block 3

Conv2D Layer
  • 128 filters, 3×3 kernel, ReLU activation
  • High-level feature extraction
  • Pneumonia-specific patterns (consolidations, infiltrates)
MaxPooling2D
  • 2×2 pool size
  • Final dimension: 28×28

​
Fully Connected Layers

Flatten
  • Converts 3D tensor to 1D vector
Dense Layer (128 neurons)
  • ReLU activation
  • Combines features for decision making
Dropout (0.5)
  • Regularization: Prevents overfitting
  • Randomly deactivates 50% of neurons during training
  • Forces network to learn robust features
Output Layer (2 neurons)
  • Softmax activation
  • 2 classes: [NORMAL, PNEUMONIA]
  • Outputs probabilities that sum to 1

​
Activation Functions

​
ReLU (Rectified Linear Unit)

Used in all hidden layers: 
f(x) = max(0, x)
Advantages of ReLU:
  • No vanishing gradient problem
  • Computationally efficient
  • Introduces non-linearity
  • Standard in modern CNNs

​
Softmax

Used in output layer: 
softmax(xi) = exp(xi) / Σ exp(xj)
Advantages of Softmax:
  • Converts logits to probabilities
  • Clear probabilistic interpretation
  • Ideal for multi-class classification

​
Hyperparameters

​
Training Configuration

Loss Function: Categorical Crossentropy
  • Penalizes incorrect predictions
  • Standard for classification tasks
Optimizer: Adam
  • Adaptive learning rate
  • Combines advantages of RMSprop and Momentum
  • Fast and stable convergence
  • Initial learning rate: 0.001
Adam is chosen over SGD for its adaptive learning rate, which helps navigate the complex loss landscape of deep networks more efficiently.
Metrics
  • Accuracy: Percentage of correct predictions
  • Precision: Of predicted pneumonia cases, how many are correct
  • Recall (Sensitivity): Of actual pneumonia cases, how many we detect
  • F1-Score: Harmonic mean of precision and recall

​
Regularization Strategies

  • Prevents co-adaptation of neurons
  • Reduces overfitting
  • Simulates ensemble of networks
  • Applied only during training
Applied during training:
  • Rotations: ±15 degrees
  • Translations: ±10%
  • Zoom: ±10%
  • Horizontal flip: Yes
Goal: Increase dataset variability and improve generalization
  • Monitors validation loss
  • Stops training if no improvement for 5 epochs
  • Prevents overfitting to training data
  • Restores best weights

​
Model Parameters

​
Parameter Count Estimation

Conv1:     32 × (3×3×3 + 1) = 896 parameters
Conv2:     64 × (3×3×32 + 1) = 18,496 parameters
Conv3:     128 × (3×3×64 + 1) = 73,856 parameters
Dense1:    128 × (128×28×28) + 128 ≈ 12.8M parameters
Output:    2 × (128 + 1) = 258 parameters

Total: ~13M parameters

​
Model Size Justification

​
Why This Architecture?

Not Too Deep
  • Avoids overfitting with limited dataset (~6K images)
  • Faster training time
Not Too Shallow
  • Sufficient capacity to learn complex patterns
  • Can capture hierarchical features
Balanced
  • Between accuracy and training time
  • Can be trained on CPU in reasonable time
  • Production-ready for deployment

​
Alternative Approaches Considered

Advantages:
  • Better potential accuracy
  • Less training data needed
  • Pre-trained on ImageNet
Disadvantages:
  • Greater complexity
  • Harder to explain
  • Larger model size
Decision: Not used to maintain simplicity and educational value
Advantages:
  • Faster training
  • Fewer parameters
  • Lower computational requirements
Disadvantages:
  • Insufficient capacity
  • Lower performance
  • Cannot capture complex patterns
Decision: Insufficient for pneumonia detection complexity

​
Progressive Feature Learning

The architecture follows a pattern of progressive feature extraction: Layer 1 (32 filters): Basic features
  • Edges and lines
  • Simple gradients
  • Basic texture elements
Layer 2 (64 filters): Mid-level features
  • Texture patterns
  • Opacity variations
  • Shape components
Layer 3 (128 filters): High-level features
  • Consolidation patterns
  • Infiltrate signatures
  • Disease-specific markers
This hierarchical feature learning is what makes CNNs so effective for medical image analysis - each layer builds upon the previous one to create increasingly sophisticated representations.

​
Technical References

  • He et al. (2016) - Deep Residual Learning
  • Simonyan & Zisserman (2014) - VGG Networks
  • Rajpurkar et al. (2017) - CheXNet: Radiologist-Level Pneumonia Detection

Build docs developers (and LLMs) love

Get started for freeTalk to us