Medical images from computed tomography, magnetic resonance, and positron emission tomography contain far more information than a radiologist can extract with the naked eye. Radiomics extracts quantitative features describing shape, texture and intensity patterns within a region of interest (e.g., a tumor) that can then be used to develop machine learning models for predicting clinical outcomes [1].

However, identifying the most predictive features for a given clinical problem remains a challenge, as large feature sets increase the risk of overfitting and reduce the interpretability of the model, particularly in small datasets [2]. Addressing these challenges often requires programming expertise and familiarity with machine learning methods, limiting the accessibility of radiomics for researchers without specialized technical backgrounds.

To improve the accessibility of radiomics, we present MEDiml (www.mediml.app ), which stands for Medical Image Analysis with Machine Learning. MEDiml is an open-source software available both as a Python package (MEDiml-pkg) and as a desktop application (MEDiml-app). MEDiml covers the full radiomics workflow, including image processing, feature extraction, model training and performance evaluation.

MEDiml is built around complexity minimization [3]. Features are grouped into complexity levels according to how much computation they require, from shape and size, to intensity, to texture, to features from filtered images. Simpler models are less prone to overfitting in small datasets, are more generalizable, and are easier for clinicians to interpret [4]. For this purpose, MEDiml trains a model at each level and compares the levels with statistical tests, such that the simplest level with sufficient predictive performance is retained. Moreover, feature extraction in MEDiml follows international standards for radiomics extraction, set by the Image Biomarker Standardization Initiative (IBSI) guidelines [5, 6], making feature extraction reproducible across studies.

Finally, clinical researchers without programming experience can design and execute radiomics experiments through a graphical interface, while other users can directly interact with the Python package or export interface-generated workflows as executable Python code.

Radiomics has a well-documented reproducibility problem, as feature values can shift substantially depending on image preprocessing choices, discretization parameters, and whether the software follows standardized definitions. A study by Zwanenburg et al. [7] found that only a subset of radiomic features remained robust under image perturbations such as noise addition, rotation and contour variation. Although clear guidelines and standards now exist, the IBSI being the most significant, clinical translation remains poor [8, 9].

The high dimensionality of radiomic feature sets is another significant challenge. Feature sets can exceed 10,000 values per scan, especially when image filters are applied [10]. However, medical imaging datasets are rarely large enough to handle the required ratio of features to patients needed to prevent overfitting in specialized predictive modeling tasks [2]. Furthermore, the predictive value of radiomic features is highly dependent on the clinical context. For example, in lung cancer survival prediction, tumor volume alone has been shown to carry most of the prognostic signal, making texture features redundant [9], whereas wavelet-transformed texture features were able to distinguish tissue subtypes in other applications [11].

Finally, the successful application of radiomics requires bridging the expertise of interdisciplinary teams: clinical researchers who understand the medical context, and computer scientists who can implement the analytical workflows. Nonetheless, tools designed primarily for one group may not meet the needs of the other.

Several radiomics tools exist, ranging from open-source feature extraction libraries to graphical platforms. Their coverage of feature extraction, graphical interaction and predictive modeling varies, as does their reported coverage of the IBSI recommendations (see Table 1 for a comparison with MEDiml).

MEDiml’s specific contribution is the combination of IBSI compliance, a zero-code graphical interface, a Python package, integrated machine learning tools, and a code-generation feature that connects the graphical and programmatic options. In many of the reviewed tools, predictive modeling is either absent or limited to isolated functionalities, requiring users to export features to external software for feature selection, model development, or statistical analysis. Moreover, none of these tools connects graphical and programmatic interfaces through code generation.

Pan/Zoom

A key architectural characteristic of MEDiml is that the graphical interface and the Python package run on the same underlying computation, rather than being independent products, to ensure methodological consistency and reproducibility across components.

The default radiomic workflow in MEDiml includes the following steps: feature extraction, feature cleaning to remove constant, quasi-constant, and duplicate features, as well as feature normalization using ComBat [20] to mitigate non-biological batch effects arising from differences in scanner manufacturers, acquisition protocols, or reconstruction settings. Feature selection is then performed using the False Discovery Avoidance (FDA) approach [21] to identify the smallest subset of non-redundant predictive features associated to the clinical outcome. The selected features are subsequently used to train machine learning models using the PyCaret framework [22], with an integrated analysis mode that explains predictive decisions using individual feature importance driving the model’s performance. This workflow is depicted in Figure 1, Block 2.

Pan/Zoom

The MEDiml Python package (MEDiml-pkg) provides the computational framework used by both programmatic analyses and the graphical application. It is compliant with IBSI 1 and IBSI 2 [5, 6], builds upon widely adopted scientific Python libraries, including NumPy [24], Pandas [25] and PyCaret [22], and uses a modular architecture where each stage of the workflow is implemented as an independent component (Figure 1, Block 3, provides a summary of the architecture):

• Wrangling: This module covers data loading through the DataManager class, where all imaging scans are loaded, along with their corresponding masks, and prepared for subsequent processing. This module also provides users with statistics, including voxel intensity distribution analyses, to assist in the selection of the appropriate radiomic feature extraction parameters.
• Processing: This module implements all methods required for data processing steps recommended by the IBSI, including interpolation, region of interest (ROI) extraction, re-segmentation, intensity discretization, and Standardized Uptake Value (SUV) computation for Positron Emission Tomography (PET) imaging, which are fundamental to radiomics analyses [5].
• Filters: This optional module applies filtering techniques on images prior to feature extraction, to reduce noise, highlight anatomical structures, and generate new image representations, either using linear filters (mean, Laplacian-of-Gaussian, Laws, Gabor and wavelets) or nonlinear textural filters based on the gray-level co-occurrence matrix (GLCM) [26], implemented in the TexturalFilter class. The latter are optimized to run on a Graphics Processing Unit (GPU) for computational efficiency [27, 28].
• Biomarkers: This core module performs the extraction of all IBSI-compliant radiomic feature families, from original or filtered images, for individual scans or entire cohorts through the BatchExtractor class. The Ray library [29] is used to parallelize extraction tasks for faster computation.
• Learning: This module encompasses the complete machine learning workflow. The Feature Set Reduction (FSR) class reduces the feature sets using the FDA method [21], and the RadiomicsLearner class trains and evaluates models across radiomic complexity levels, using any PyCaret algorithm with \( k\) -fold cross-validation [30], random subsampling or bootstrapping. The Results class then compares complexity levels using statistical significance tests (DeLong’s [31, 32], Wilcoxon [33], paired [34] and Nadeau-Bengio corrected t-tests [35]) and performance heatmaps, and interprets models through feature-importance histograms and Explanation Trees, which trace predictive contributions from the complexity level to filters, feature families and individual features. Finally, the optimal pipeline is retrained on the entire learning set, evaluated on a holdout set if provided, and exported for deployment.
• Utils: This module contains general utility and shared components that support the operation of the other software modules.

The MEDiml desktop application (MEDiml-app) is a zero-code graphical interface in which every action invokes the same underlying Python package methods (Figure 2A). The application is built using Electron [36], which packages a web-based interface as a native desktop application. Its drag-and-drop, node-based workflow editor is implemented using React Flow [37] (https://reactflow.dev ), a package for constructing graph-based editors, and is organized around three hierarchical concepts. A node represents a specific processing step, such as data loading, feature cleaning, or model training, with configurable parameters and well-defined inputs and outputs. A pipeline is an ordered sequence of connected nodes, making the structure of the experiment explicit. A scene contains one or more pipelines that can be executed and compared simultaneously, and is the unit for saving, sharing and reproducing analyses.

The application exposes the package through four modules that follow the stages of a radiomics study:

• Single-Scan Extraction: extracts features from an individual image and its mask to explore and validate processing and extraction settings.
• DataManager: provides graphical access to the DataManager class to register a dataset and run exploratory analyses that guide the choice of extraction settings.
• BatchExtractor: runs parallelized extraction across the full cohort, with settings configured per imaging modality, and produces a single feature table that can be split by feature family (morphological, intensity and texture) through its feature categorization step.
• Learning: provides nodes for each stage of predictive modeling (Split, Design, Data, Cleaning, Normalization, Feature Reduction, Radiomics Learner and Analyze), allowing users to train, compare and interpret models across radiomic complexity levels, datasets or imaging modalities within the same scene.

The full MEDiml architecture is thoroughly explained in a published reference manual [38].

4.1 Code Generation

MEDiml has been applied in two radiomics studies [3, 39], in which the software handled image processing, feature extraction, predictive modeling, and analyses of results. In the first study, MEDiml was applied to five cancer imaging datasets, totaling 2,104 scans; the software successfully identified, for each task, the simplest complexity level achieving optimal predictive performance, focusing on less complex models. The associated results are presented in Table 2.

These studies demonstrated the utility of MEDiml’s complexity-minimization framework, emphasizing simpler and interpretable models. For example, in MRI-based RCC subtype classification, texture features were selected as the optimal complexity level (Figure 3A), and the Explanation Tree traced their contribution down to individual features (Figure 3B).

Pan/Zoom

MEDiml started as a tool primarily to support the core developer’s own research, but it has since expanded to serve interdisciplinary teams. The package has been downloaded over 500 times from the Python Package Index (PyPI) to this date, reflecting a growing user base beyond its origins.

To support users, all information about the software is available on the official website: https://mediml.app . Overall, MEDiml provides the following set of resources:

1. For the Python package:
1. Package: PyPI .
2. Documentation: ReadTheDocs .
3. Tutorials: Jupyter Notebooks .
4. Source Code: GitHub .
2. For the desktop application:
1. Application: Download Page .
2. Documentation on GitBook .
3. Video tutorials: YouTube .
4. Source Code: GitHub .
3. A community Discord server .

MEDiml is maintained and actively improved by the MEDomicsLab research group (https://www.medomicslab.com ) at McGill University.

We thank Guillaume Blain, Sarah Denis, and Andréanne Allaire for their valuable contributions to the development and enhancement of MEDiml’s user interface. We also thank Corentin Gauthier, Clarisse Cheng, and Anthonin Falk for their valuable contributions to the development of the MEDiml package. Finally, we gratefully acknowledge Martin Lepage for co-supervising a significant portion of this work [3].

Martin Vallières acknowledges the support of Unité de soutien au Système de Santé Apprenant (SSA) Québec.