hectorjg/final_project

Project: Image classification of x-ray pneumonia

The goal is to predict as accurately as possible if the chest x-ray image shows pneumonia or if it is normal. The model is a CNN of deep learning. The dataset used for this project can be found on www.kaggle.com/datasets/paultimothymooney/chest-xray-pneumonia/data

Objectives

• Understand the information (image size, colors, size of the dataset, classes).
• Build the architecture of the model considering the nature of the information.
• Understand the plots to improve the model.
• Test the model and save it.

Project Steps:

Data pre process

• Load the dataset and select the colors (Gray scale or RGB).
• Resize the images.
• Define the callbacks.

Data Analysis

• Check for imbalance in the data.

Processing Data

• Data augmentation.

Building the model

• Using Sequential() as a model.
• 1 Conv2D(32), 2 Conv2D(64), 1 Conv2D(128), 1 Convd2D(256), Flatten, 2 Dense (128, 64), Output.
• Choose the Output Activation.

Compile to train the model

-Select the optimizer.

-Choose the Loss metrics.

-Choose the accuracy metrics.

Evaluate Accuracy

• Accuracy.
• F1-score.
• Precision.
• Recall.
• Plot of the Loss in train and val.
• Model.predict() on test set.

Conclussion:

-The Learning rate of the model is slow, meaning that the model has problems converging.

-Even if the accuracy archives a 92.30% accuracy, it might need a deeper architecture to perform better.

---

Model: cnn_92.joblib

Aditional dataset: Diabetic retinopathy

• This is a CNN model for a multiclass classification problem with 5 different classes.
• This dataset has all the images in the same size, but the number images per class is imbalanced.
• The architecture may be similar to the architecture of "pneumonia", but all the parameters are different due to the nature of the images.
• The dataset can be find in: www.kaggle.com/datasets/sachinkumar413/diabetic-retinopathy-preprocessed-dataset.