Learning
Our software provides tools for model training using extracted features and facilitates the analysis of machine learning results, with a focus on identifying the optimal feature types.
Last updated
Our software provides tools for model training using extracted features and facilitates the analysis of machine learning results, with a focus on identifying the optimal feature types.
Last updated
A tutorial video is available at the bottom of the page
Like feature extraction, the learning module also utilizes a drag-and-drop design, allowing users to define their machine learning pipeline which incorporates feature processing steps such as cleaning, feature selection, and more.
The next sections explain how to use this module to train your models and analyze results.
The following image depicts the available resources in the leaning module interface:
Having gained an understanding of the various components of the extraction interface, the following section illustrates the available nodes, highlighting their individual functionalities, inputs, and outputs:
Based on the chosen options, this node generates and stores a holdout set for subsequent model evaluation.
No inputs
Design
The design node allows you to configure various options for your machine learning experiment, including the experiment name, data splitting method, train and test portions, and more. Additionally, this node will create a separate folder to store all files related to the experiment.
Split
Data
This is employed to specify your feature set. After setting the features folder, you can choose the features file (CSV files) to use in your experiment.
Design
Cleaning
Normalization
Feature Reduction
Radiomics Learner
Data cleaning consists of four parts:
Remove missing features
Remove missing patients
Remove invariant features
Impute missing features
Data
Normalization
Feature Reduction
Radiomics Learner
Normalizing radiomics features using ComBat method to mitigate batch effects arising from data from different sources.
Data
Cleaning
Feature Reduction
Radiomics Learner
Data
Cleaning
Normalization
Radiomics Learner
Data
Cleaning
Normalization
Feature Reduction
Analyze
Radiomics Learner
No outputs
A detailed demonstration on renal cell carcinoma dataset can be found in the next page.
Feature selection consists of selecting predictive features using the false discovery avoidance method. All the parameters are explained in the method's .
This node sets all options to the model used. is the only algorithm supported for now, and its settings can be tuned using an automatic approach: or using a random or a grid search.
The final node is used for analyzing the results obtained from the experiment. You can utilize this node to select specific analysis methods, such as histograms, heatmap, or importance tree. Each method comes with its own set of options that need to be configured. For additional information, please refer to the available in MEDimage.
