Welcome to our Github repository for our exam project in the First Year Project course. The project is a collaboration between adaro, aksv, bhei, swae - Group 7.
The project focuses on segmentation, feature extraction and classification of medical images. In particular images of skin lesions. The goal of the project is to train a classifier, which is able to predict the probability that an unknown skin lesions is cancerous.
All of the extracted features can be found in the csv file data/features.csv. These features were created by running the python file _1features.csv. The file contains the extracted features for each image.
By using the metadata and the extracted features, it is possible to train a series of classifiers. We chose to use KNN and Logistic Regression as classifiers. Training of the classifier is done using the file 2_train_classifiers.py. Furthermore, evaluation metrics are found in the same file. The evaluation of the classifiers is done by calculating F1-score, ROC AUC-score, Accuracy score, and plotting a confusion matrix. To simulate these metrics with most confidence, we simulate different train test splits. The simulation is run 200 times and chooses different sets of the train and validation data. The test data is the same subset of skin lesions every time. These test runs are plotted using the one_out_boxplots function.
To evaluate our chosen classifier, use the _3evaluate_classifier.py file. The file contains a function classify which takes an unknown image and mask as input and outputs the probability of the masked image being cancerous. The classify function is provided with the correct classifier saved in the root directory as a .sav file. classify imports the function features2Dataframe which extracts features from the provided image.
The segmentation was done using more simple methods. These are found in Segmentation.py. We also segmented the images using Segment Anything Model (SAM), created by META AI. We use these masks to train a Double U-net for which the weights can be found here.
The image data along with the metadata used for the project is provided by Mendeley Data and can be found here