Medical Image Analysis with Machine Learning (MEDiml): Strategies for Simpler Radiomics
Abstract
Purpose
Clinical translation of radiomics is hindered by the computational complexity of the feature extraction process, high-dimensional feature sets, and lack of accessible tools for healthcare professionals. We introduce MEDiml, a new open-source platform designed to help democratize the development of radiomics models, by identifying the simplest predictive feature sets through both code-based and no-code graphical interface.
Methods
MEDiml was evaluated using 89,714 features from five oncological datasets (n = 2,104). Tasks included histology subtype prediction for non-small cell lung cancer (NSCLC, n = 506, MRI) and renal cell carcinoma (RCC, n = 599, MRI); RCC (n = 326, contrast-enhanced CT); IDH1 mutation prediction for low grade glioma (LGG, n = 344, MRI); and grade prediction for meningioma (n = 329, MRI). Features were stratified by computational complexity levels: morphological, intensity, texture, linear filters, and nonlinear filters. MEDiml isolates the simplest and most informative features by removing unstable and redundant features.
Results
We successfully created MEDiml, an open-source software, provided with a one-click installer, a comprehensive documentation, that can all be explored via mediml.app. Evaluation across the five datasets demonstrated that maximum predictive performance does not always require high-complexity features. Optimal complexity levels were: morphological for LGG IDH1 and Meningioma grade; intensity for NSCLC and RCC (CECT); and texture for RCC (MRI). In the RCC CECT cohort, optimizing the re-segmentation range, a parameter for excluding non-target voxels that affects intensity features, improved performance from an AUC of 0.82 to 0.86.
Conclusion
MEDiml successfully identifies the "optimal complexity level" for specific clinical outcomes, demonstrating that simpler, more interpretable models can often match or exceed the performance of high-dimensional sets. By providing a user-friendly interface and a strategy for feature minimization, MEDiml lowers the barrier to entry for clinicians and provides a scalable pathway for the clinical integration of radiomic biomarkers.