Chest X-Ray 20-Class Prediction using Deep Learning

Diagnostic Classes

224×224

Image Resolution

EfficientNet-B0

Transfer Learning Model

20 Labels

Expected Accuracy

Project Overview

This project uses the NIH Chest X-ray dataset containing 51,043 training images and 20 diagnostic classes, including Atelectasis, Cardiomegaly, Pneumonia, Pneumothorax, and No Finding. Images are resized to 224×224, augmented, and used to fine-tune EfficientNet-B0 with a custom classification head. The model outputs one of 20 possible thoracic diagnoses.

Diagnostic Examples

Representative visualizations highlighting the contrast between healthy lungs and pneumonia-affected regions.

Pipeline Architecture

X-Ray Images

Preprocessing

Augmentation

EfficientNet-B0

20-Class Prediction

Key Code Snippets

Image Preprocessing

transform = transforms.Compose([
    transforms.Resize((224, 224)),
    transforms.ToTensor(),
    transforms.Normalize(mean=[0.485], std=[0.229])
])

Data Augmentation

train_transform = transforms.Compose([
    transforms.RandomHorizontalFlip(),
    transforms.RandomRotation(10),
    transforms.Resize((224, 224)),
    transforms.ToTensor()
])

Model Definition

base_model = EfficientNetB0(weights="imagenet", include_top=False,
    input_shape=(224, 224, 3))

x = GlobalAveragePooling2D()(base_model.output)
x = Dropout(0.4)(x)
outputs = Dense(20, activation="softmax")(x)
model = Model(base_model.input, outputs)

Training Loop

for images, labels in train_loader:
    outputs = model(images)
    loss = criterion(outputs, labels)
    loss.backward()
    optimizer.step()

Prediction

with torch.no_grad():
    probs = torch.softmax(model(x), dim=1)
    pred = torch.argmax(probs, dim=1)

Model Performance

Evaluation Metrics

Accuracy: Classification accuracy across 20 thoracic conditions
Precision: Positive prediction reliability
Recall: Pneumonia detection sensitivity
F1 Score: Precision-recall balance
ROC-AUC: Threshold-independent performance

Dataset Summary

Training Samples: 51,043

Test Samples: 17,015

Classes: 20 thoracic conditions

Input Shape: 224 × 224 × 3

Included labels: Atelectasis, Cardiomegaly, Consolidation, Edema, Effusion, Emphysema, Fibrosis, Hernia, Infiltration, Mass, Nodule, Pleural Thickening, Pneumonia, Pneumothorax, Pneumoperitoneum, Pneumomediastinum, Subcutaneous Emphysema, Tortuous Aorta, Calcification of the Aorta, and No Finding.

Technologies Used

Python PyTorch Torchvision OpenCV NumPy Matplotlib

Real Images from the Notebook

These are actual figures extracted directly from the Jupyter notebook, including dataset samples, plots, and model visualizations.