๐ป Image Classification: Detecting Cancer Metastases on Pathology Images
This repository is dedicated to a project revolving around the early detection of metastatic tissue in histopathologic scans of lymph node sections. This is paramount for cancer prognosis and the determination of treatment methods.
The project encompasses the following steps:
- ๐ป Data Preprocessing: Cleaning and preprocessing the dataset for optimal performance in the subsequent modeling steps.
- ๐ป Data Augmentation: Leveraging techniques to synthetically enlarge the dataset, ensuring a robust model training.
- ๐ป Model Building: Setting up different model architectures, including a basic CNN model and a model inspired by the MNIST dataset.
- ๐ป Hyperparameter Tuning: Using techniques like KerasTuner to find the most optimal hyperparameters for our models.
- ๐ป Model Evaluation: Analyzing performance metrics, and drawing conclusions from the trained models.
- ๐ป Predictions: Using the best-performing model to make predictions on test data.
This project utilizes the following technologies:
- Python
- TensorFlow/Keras
- KerasTuner
These tools and libraries are essential for data processing, model training, and evaluation.
To execute the project on your local machine:
- Ensure all data is installed using
kaggle competitions download -c histopathologic-cancer-detection
. - Run the Jupyter notebook to walk through the data processing, model training, and prediction stages.
The projected is structured as follows:
โโโ train
โ โโโ images
โ โโโ labels
โโโ test
โ โโโ images
โโโ models
โ โโโ basic_cnn
โ โโโ MNIST_inspired
โ โโโ advanced_model
โโโ tuning_results.csv
โโโ README.md
train/images
: Directory with training images.train/labels
: Associated labels for the training images.test/images
: Directory with test images.models
: Contains saved models from the project.tuning_results.csv
: Results from hyperparameter tuning phase.
The trained models and their evaluations are available within the Jupyter notebook. The most advanced model reached an accuracy of ~91.13% on the validation set, showcasing its potential in the medical imaging field.
A shoutout to the Kaggle community for providing such a rich dataset and for fostering a competitive environment that drives innovation. This particular dataset and competition were invaluable resources for this project.
This endeavor is protected under the MIT License. This grants the liberty to use, adapt, and distribute the content for both academic and commercial applications.
For additional details or inquiries, feel free to drop an email at [email protected].