MEDical Imageanalysis withMachine Learning

MEDiml combines a Python package (MEDiml-pkg) and a graphical desktop application (MEDiml-app) for radiomics workflows. The two interfaces share a computational engine, and users can export workflows built in the desktop interface as Python code. Feature extraction follows IBSI standards. The software compares models built from progressively more complex feature groups and selects the simplest group whose performance is not significantly surpassed. Across five oncological datasets containing 2,104 scans, reported AUCs ranged from 0.65 to 0.82; shape or intensity features were sufficient for four tasks.

This reference manual presents MEDiml, an open-source radiomics platform combining a Python package (MEDiml-pkg) and a graphical desktop application (MEDiml-app). Both interfaces share the same computational implementation, with graphical workflows exportable as Python code. The manual covers radiomics concepts, image preprocessing, feature extraction, predictive modeling and interpretation, alongside the software architecture and installation. It compares MEDiml with existing tools and illustrates how different levels of radiomic complexity affect predictive performance across five oncological imaging cohorts comprising 2,104 scans.

The high dimensionality of radiomic feature sets and potentially high computational requirements underscore the need for an effective method to identify the smallest set of predictive features. This work develops a methodology and tools (MEDiml) to identify and explain the smallest set of predictive radiomic features across five cancer datasets (n=2,104), estimating the optimal radiomics complexity level for specific medical outcomes.

Datasets

5 cancer cohorts

Patients

2,104

Features

89,714 extracted

Subject

eess.IV / cs.CV