How joberthcastro/ia-diagnostico-pneumonia works
Project Type and Purpose
• This is a standalone machine-learning image classification pipeline implemented in Python (no separate frontend/backend).
• Its goal is to train, validate, test, and evaluate a convolutional neural network (CNN) for binary classification of chest X-ray images (normal vs. viral pneumonia).
File Structure Analysis
• Dataset/ contains three subfolders (Train, Val, Test), each split into Normal_Radiografia and Viral_Radiografia.
• The single script pneumonia_classifier.py orchestrates data loading, preprocessing, model definition, training, evaluation, and artifact generation.
• requirements.txt lists the Python dependencies (TensorFlow/Keras, OpenCV, NumPy, Matplotlib, Seaborn, scikit-learn, pandas).
Architecture Insights from README
• A CNN with 3 convolutional+ReLU+MaxPooling layers, dense layers with dropout, and sigmoid output for binary classification.
• Preprocessing pipeline: grayscale conversion → histogram equalization → Gaussian blur → normalization → resizing to 225×225.
• Training: 15 epochs, data augmentation, early stopping, best model saved (best_model.h5).
• Outputs: training history plot, confusion matrix, per-class metrics chart, ROC curve.
Main Components
a. Data Ingestion & Splitting
• Raw images (Train/, Val/, Test/)
b. Preprocessing Pipeline
• Grayscale → Equalization → Denoising → Normalization → Resize
c. Data Augmentation Module
• On-the-fly random transformations (rotation, shift, flip, etc.)
d. Model Definition
• Sequential CNN layers + dense layers + dropout → sigmoid output
e. Training Engine
• Epoch loop, optimizer, loss function, callbacks (EarlyStopping, ModelCheckpoint)
f. Evaluation & Metrics
• Predict on test set → compute accuracy, precision/recall/F1, confusion matrix, ROC/AUC
g. Artifacts & Outputs
• Saved model (best_model.h5), plots (training_history.png, confusion_matrix.png, classification_metrics.png, roc_curve.png)
Relationships and Data Flow
• Preprocessed data feeds into Model Definition & Training Engine
• Training Engine outputs checkpoints to Artifact store
• After training, Test data flows through Preprocessing → Model → Evaluation module
• Evaluation produces metrics and plots
Architectural Patterns & Design Principles
• Modular pipeline: separate concerns (I/O, preprocessing, augmentation, modeling, evaluation)
• Callback pattern in training (early stopping, checkpointing)
• Batch processing for efficiency
• Reproducibility via fixed random seeds and saved model
Technologies & Libraries
• TensorFlow/Keras for model definition/training
• OpenCV for image preprocessing
• NumPy/pandas for data manipulation
• scikit-learn for metrics
• Matplotlib/Seaborn for plotting