Reference Manual of MEDiml

1 Introduction

• Radiomics extracts quantitative image features to characterize tissues and may support diagnosis, prognosis, and more precise personalized treatment.The manual uses “radiomics” specifically for handcrafted radiomics and describes IBSI nomenclature, definitions, workflows, and reference data for standardization and verification.Source: §1 · pp. 4–5
• Radiomic features include morphology, intensity distributions, texture, and features calculated after filters modify image intensities.Morphological features describe region shape and size; intensity features describe pixel-intensity characteristics; texture features summarize spatial intensity patterns.Source: §1 · pp. 4–5
• Existing tools often require programming experience and serve either clinical research or computational work, rather than both.MEDiml is presented as an open-source, IBSI-compliant platform intended to improve accessibility and facilitate multidisciplinary collaboration.Source: §1 · p. 5
• MEDiml combines a Python package and graphical application on the same computational framework, with code generation exposing visual workflows as inspectable, extendable Python code.Its Explanation Tree organizes model-specific feature contributions hierarchically, alongside statistical comparisons and feature-importance histograms.Source: §1 · p. 5
• MEDiml compares models using progressively more complex radiomic feature sets to identify compact sets suited to a clinical question.The feature categories include morphological, intensity, texture, and filter-derived features.Source: §1 · pp. 4–5

2 Resources

• The official website is the main entry point for MEDiml capabilities, downloads, documentation, and tutorials.Resources are organized for the Python package and desktop application.Source: §2 · p. 5

2.1 Python Package

• The Python package resources include PyPI distribution, API documentation, and a source repository with issue tracking.The documentation covers installation, input organization, configuration files, and the API.Source: §2.1 · pp. 5–6

2.2 Desktop Application

• The desktop application resources provide Windows, macOS, and Linux installers, application documentation, workflow tutorials, and source-code issue tracking.The documentation covers installation, graphical workflow construction, modules, troubleshooting, and contributions; a demonstration uses soft-tissue sarcoma data.Source: §2.2 · p. 6

2.3 Community and Development

• A Discord channel provides shared discussion, while MEDomicsLab’s website describes the research group and its activities.For reproducible problems, the affected component’s issue tracker provides a record that can be followed during development.Source: §2.3 · p. 6

3 Background: Radiomics

• Radiomics extracts quantitative image features to characterize tissues and uses them to train models for diagnosis, classification, or outcome prediction.The IBSI provides definitions and reference values for 173 features.Source: §3 · p. 6

3.1 Image Processing

• MEDiml’s processing subpackage handles image operations, while segmentation masks identify ROI voxels and restrict feature extraction to that anatomical region.A binary mask uses 1 for voxels within the ROI and 0 for voxels outside it.Source: §3.1.1 · p. 6, §3.1 · p. 6

3.1.2 Interpolation

• Interpolation supports consistent isotropic voxel spacing; re-segmentation excludes voxel ranges from an ROI, and ROI extraction marks outside-mask voxels as NaN.IBSI recommends trilinear interpolation as a conservative default, while MEDiml supports multiple methods.Source: §3.1.2 · p. 7, §3.1.3 · p. 7, §3.1.4 · p. 7

3.1.5 Discretization

• Discretization makes texture-feature calculation tractable; MEDiml implements Fixed Bin Number, Fixed Bin Width, and intensity histogram equalization.The first two approaches use a user-defined bin count or bin width, respectively.Source: §3.1.5 · p. 7

3.2.1 Morphological features

• MEDiml extracts feature families through application modules or its package biomarkers subpackage; morphological features describe ROI geometry.It computes 27 morphological features by default and excludes Moran’s I and Geary’s C unless users enable them, because they are computationally expensive.Source: §3.2 · p. 7, §3.2.1 · p. 7

3.2.2 Intensity features

• Intensity features describe voxel-intensity distributions, including local, statistical, histogram, and intensity-volume histogram families.Local features use neighborhoods around ROI voxels, while histogram-based families use discretized intensities.Source: §3.2.2 · pp. 7–8

3.2.3 Texture features

• Texture features summarize spatial relationships among discretized gray levels through GLCM, GLRLM, GLSZM, GLDZM, NGTDM, and NGLDM matrices.These families count neighboring pairs, runs, connected zones, edge distances, neighborhood gray-level differences, or gray-level dependencies.Source: §3.2.3 · p. 8

3.2.4 Filter-based features

• After filtering alters image intensities, MEDiml calculates intensity-histogram and texture features rather than all feature types.The modified intensities may no longer retain their original physical meaning.Source: §3.2.4 · p. 8

4.1 Radiomics Feature Extraction

• Feature extraction preprocesses medical images, applies category-specific settings, and uses parallelized batch processing to produce scan-level JSON files and a dataset-level CSV.The CSV is formatted for the subsequent machine-learning step.Source: §4.1 · p. 9

4.2 Modeling

• Predictive modeling curates features by removing invariant, excessively missing, or highly correlated variables, then normalizes their scales before training.These steps prepare the extracted radiomic data as model inputs.Source: §4.2 · p. 9

4.2.1 Feature Set Reduction (FSR)

• Feature set reduction addresses high-dimensional datasets by selecting task-relevant features and removing redundant, irrelevant, or noisy variables.Feature selection can improve interpretability and reduce computational complexity, and often enhances predictive performance.Source: §4.2.1 · pp. 9, 11

4 MEDiml radiomics workflow

• MEDiml organizes radiomic features by complexity and compares models across levels to select the simplest level with the highest statistically significant performance.The selected level balances feature count and complexity against predictive performance.Source: Figure 2 · p. 10, §4.2.2 · p. 11

4.2.2 Training and Testing

• Models are trained and evaluated at progressively higher radiomic complexity levels, with statistical testing used to identify the simplest level achieving the best predictive performance.The workflow supports cross-validation or random subsampling and retrains models across iterations to select parameters.Source: §4.2.2 · p. 11

4.3 Results Analysis and Interpretation

• Results analysis compares performance across complexity levels with statistical tests, selecting the lowest-complexity model with statistically equivalent or superior performance and no detected significant gain at higher complexity.MEDiml supports DeLong’s test for AUC comparisons and paired t-test and Wilcoxon signed-rank test options.Source: §4.3 · pp. 11–12

5.1 Overview

• MEDiml-pkg supports IBSI phases 1 and 2 and parallelized feature extraction, emphasizing compact, informative feature subsets that maintain strong predictive performance and interpretability.The package’s workflow connects loading and processing to extraction, feature reduction, modeling, comparison, and interpretation.Source: §5.1 · p. 12

5.2 Data Wrangling

• DataManager loads DICOM and NIfTI images, associates scans with segmentation masks, extracts metadata, and provides exploratory intensity analyses for quality control and parameter selection.Support for NRRD is planned for future releases.Source: §5.2 · p. 12

5.3 Image Processing

• The processing subpackage implements IBSI-recommended preprocessing, including interpolation and resampling, ROI extraction, re-segmentation, intensity discretization, and PET SUV computation.Users can choose among processing options, including nearest-neighbor, linear, spline, and cubic interpolation.Source: §5.3 · pp. 12–13

5.4 Filtering

• Filtering is optional and can enhance image characteristics, reduce noise, or generate derived representations that may contain additional predictive information.MEDiml supports five filter classes, including Laplacian-of-Gaussian filtering.Source: §5.4 · p. 13

5.4.1 Nonlinear Filtering

• Textural filtering iteratively computes local GLCM texture features with a three-dimensional cubic kernel, producing 25 maps for use in subsequent feature extraction.The GPU-accelerated TexturalFilter currently supports GLCM filtering only.Source: §5.4.1 · p. 13

5.5 Biomarker Extraction

• BatchExtractor supports single-patient and cohort-level extraction, parallelized with Ray, across major IBSI-compliant feature families and from original or filtered images.Supported families include morphological, intensity-based, and texture features, yielding biomarkers across multiple complexity levels.Source: §5.5 · p. 13

5.6 Feature Set Reduction: The False Discovery Avoidance (FDA) method

• FDA ranks candidate features by outcome correlation and stability across repeated stratified subsamples, computing minimum and mean Spearman correlations for each feature.The default is 100 training subsamples, and ranking occurs at each complexity level and training fold.Source: §5.6 · p. 14
• Features with |Rmin| > 0.5 are retained, variants are selected by highest Rmean, and one of two features with |intracorrelation| > 0.7 is removed based on weaker outcome association.The resulting subsets typically contain 5–20 biomarkers.Source: §5.6 · p. 14

5.7 Modeling

• The modeling subpackage supports dataset partitioning, cleaning, normalization, feature reduction, training, and evaluation with PyCaret algorithms and k-fold cross-validation or random subsampling.Its central goal is to assess whether increasing feature count improves predictive performance.Source: §5.7 · p. 14

5.8 Results Analysis and Interpretation

• Statistical tests compare models across complexity levels, and performance heatmaps summarize metrics and significance to identify strong performance with minimum feature count and complexity.Supported tests include DeLong’s test for AUC, Wilcoxon signed-rank tests, paired t-tests, and the Nadeau-Bengio corrected t-test.Source: §5.8 · p. 14, Figure 4 · p. 16
• The Explanation Tree traces contributions from complexity level through filters and feature families to individual features, with branch widths proportional to importance.Feature-importance histograms provide a complementary view of relative contributions.Source: §5.8 · pp. 14, 16, Figure 5 · p. 16

5.9 Model Finalization

• Model finalization retrains the selected pipeline on the full learning dataset, serializes the model and methodological settings, and can evaluate it independently on a holdout set.A holdout evaluation can estimate performance on previously unseen data.Source: §5.9 · pp. 16–17

6 The MEDiml application

• The desktop application uses the MEDiml Python package and lets researchers conduct the radiomics workflow without programming expertise.Analyses are assembled visually from connected operations.Source: §6 · p. 17, §6.1 · p. 17

6.1 Overview

• The drag-and-drop, node-based application invokes the same underlying Python methods as the package, enabling visually designed experiments to transfer into reproducible Python workflows.The shared computational implementation supports collaboration between clinical researchers and computational scientists.Source: §6.1 · p. 17

6.2 Nodes, Pipelines and Scenes

• Nodes configure individual operations, pipelines connect nodes into ordered workflows for distinct experimental settings, and scenes contain pipelines for designing and comparing experiments.Scenes are the primary unit for saving, sharing, exporting, analyzing, and reproducing analyses.Source: §6.2 · pp. 18–19

6.3 Modules

• The application’s modules follow analysis stages from single-scan inspection through dataset preparation and cohort extraction to model training, comparison, and interpretation.The modules are Single-Scan Extraction, DataManager, BatchExtractor, and Learning.Source: §6.1 · p. 17, §6.3 · p. 19

6.3.1 Single-Scan Extraction

• Single-Scan Extraction lets users configure preprocessing and extraction on an individual DICOM or NIfTI image and inspect feature values before cohort analysis.It supports testing processing choices and validating extraction settings on reference cases.Source: §6.3.1 · p. 19

6.3.2 DataManager

• DataManager registers imaging studies and masks, extracts metadata, and provides intensity statistics and histograms to guide preprocessing and extraction settings.Its Process data step loads, reads, and organizes input data.Source: §6.3.2 · pp. 19–20, Figure 12 · p. 20

6.3.3 BatchExtractor

• BatchExtractor configures cohort-level settings per imaging modality, runs parallelized extraction, and creates a dataset CSV compatible with Learning.Feature Categorization splits the extracted table into family-specific subsets, placing unmatched features in an “unknown” set rather than discarding them.Source: §6.3.3 · p. 21, Figure 13 · p. 21, Figure 14 · p. 21

6.3.4 Learning

• Learning provides nodes for data splitting, design, data loading, cleaning, normalization, feature reduction, model training, and analysis.Multiple pipelines can be run in one scene to compare complexity levels, datasets, or imaging modalities; results are interactive and exportable.Source: §6.3.4 · p. 22

7.1 Architecture Overview

• A shared computational backend keeps graphical and programmatic workflows methodologically consistent and supports reproducibility and collaboration across clinical and computational teams.Clinical researchers can design experiments visually while computational researchers access, reproduce, and extend those workflows programmatically.Source: §7.1 · p. 22

7.2 Technology Stack

• Electron packages the desktop application for Windows, macOS, and Linux, while React and ReactFlow support interactive, node-based workflow construction.The workflow nodes represent tasks such as preprocessing, feature extraction, selection, machine learning, and result analysis.Source: §7.2 · pp. 22–23
• Go translates user-defined workflows into structured instructions, coordinates Python execution, and returns outputs; the Python package uses scientific-computing and machine-learning libraries.The package includes NumPy, Pandas, SciPy, Scikit-learn, PyCaret, and Ray.Source: §7.2 · p. 23

7.3 Code Generation

• MEDiml translates graphically configured analysis settings into Python execution and can export each workflow as an executable Jupyter Notebook.The NotebookBuilder traverses the workflow graph and converts its nodes and configuration into code that users can inspect, modify, extend, and rerun.Source: §7.3 · pp. 23–24, Figure 15 · p. 23

7.4 GPU-Accelerated Filtering

• GPU-accelerated texture filtering uses C++ and CUDA kernels to parallelize voxel-level computations, reducing processing times compared with CPU implementations.PyCUDA integrates kernel compilation, memory transfer, execution, and error handling with the Python framework.Source: §7.4 · p. 24

7.5 Dosiomics Support

• MEDiml processes radiotherapy dose maps through its shared imaging workflow and adds dose-specific biomarkers alongside standardized radiomic features.Supported dose descriptors include DVH metrics, treatment-planning indices, Vx metrics, and isocenter-related parameters.Source: §7.5 · p. 24

7.6 IBSI Compliance

• MEDiml participated in IBSI chapter 1 and 2 validation using reference datasets and phantom objects, supporting standardized and comparable radiomic feature computation.The process included preprocessing, filtering, and feature extraction, with resulting values submitted for comparison against other implementations.Source: §7.6 · p. 24

7.7 How to install MEDiml

• The package is available through PyPI, including version-specific and less-stable development-branch installations, while desktop releases are installed for Windows, macOS, or Linux.The package requires Python 3.8 or later; the desktop installer installs the application requirements during setup.Source: §7.7.1 · p. 25, §7.7.2 · p. 25

8 MEDiml Vs. Existing Radiomics Tools

• Many radiomics tools prioritize feature extraction, graphical usability, or predictive modeling rather than all three, and researchers may need multiple packages to complete a study.Predictive modeling is often absent or limited, requiring feature export to external software.Source: §8 · p. 25
• MEDiml combines an IBSI-compliant radiomics engine, integrated machine learning, a graphical interface, and a Python package in one framework.Its automatic code generation connects visually designed workflows to Python for reproduction and extension without reimplementation.Source: §8 · p. 25
• Table 1 compares tools by interface, programming language, IBSI compliance, radiomics and machine-learning capabilities, execution environment, code generation, and source-code availability.Its check marks, crosses, and “Partial” labels indicate capability availability and limited IBSI compliance, respectively.Source: Table 1 · p. 26

9.1 Cohorts

• Five cohorts cover NSCLC, LGG, meningioma, and two RCC imaging cohorts, with prediction tasks spanning histological subtype, tumor grade, and IDH1 mutation status.The cohorts combine CT/CECT and MRI data from published, public, institutional, and hospital sources.Source: §9 · p. 26, §9.1 · p. 26, Table 2 · p. 27

9.2 Radiomic Features and Complexity Levels

• The evaluation compares nine progressively combined levels, from morphological, intensity, and texture features to linear-filter and textural-filter features.Feature sets were derived from original and filtered images to assess the contribution of more elaborate extraction operations.Source: §9.2 · p. 27

9.3 Statistical analysis

• Models were evaluated with aggregated AUC across ten cross-validation test folds, and DeLong tests treated p-values below .05 as significant.Aggregating predictions into one ROC curve was used to address small-dataset issues, while sacrificing model variance.Source: §9.3 · p. 27

9.4.1 Optimal complexity level identification and explanation

• Morphological features were selected as optimal for the Meningioma cohort and LGG IDH1 prediction, while intensity features were optimal for NSCLC histological subtype and clear/non-clear RCC classification based on CECT.AUCs were 0.65 for Meningioma, 0.68 for LGG IDH1 prediction, 0.76 for NSCLC, and 0.82 for CECT-based RCC classification.Source: §9.4.1 · p. 28
• For MRI-T2WI RCC subtype classification, texture improved performance over morphological (p=.01) and intensity (p=.03) models, achieving AUC 0.72.GLCM features had the highest mean importance, with cluster shade ranking highest.Source: §9.4.1 · p. 28
• Linear-filter models were never selected as optimal, while textural-filter models were also not selected despite high AUC values across cohorts.For NSCLC subtype classification, the linear-filter model achieved AUC 0.77; the textural-filter model matched the selected level at AUC 0.76 but was not chosen because of its higher complexity.Source: §9.4.1 · p. 28, Figure 16 · p. 29

9.4.2 In-depth analysis of an optimal complexity level

• A 0.04 AUC gain, from 0.82 to 0.86, followed optimization of the ROI re-segmentation range in clear versus non-clear RCC classification using CECT.The identified optimal level was intensity-based, with median intensity the highest-importance feature.Source: §9.4.2 · p. 28, Figure 18 · p. 31

10.1 The MEDiml software

• MEDiml combines graphical and Python workflows with complexity-level comparisons to support compact, interpretable signatures rather than ever-larger feature sets.Its broader software ecosystem combines IBSI-compliant extraction, integrated machine learning, local execution, workflow editing, and automatic code generation.Source: §10.1 · p. 32

10.2 Use Cases: MEDiml Evaluation

• Across five datasets, selected complexity levels aligned with findings from other studies; in CECT-based RCC classification, 15 intensity features among 18,112 extracted features achieved the best performance.Optimizing the ROI Hounsfield unit range then improved AUC from 0.82 to 0.86.Source: §10.2 · pp. 32–33
• Limited access to large datasets constrained assessment across populations, imaging protocols, and clinical settings; the evaluation also used binary tasks and pre-existing manual segmentations.The authors call for further validation, including comparisons with deep radiomics and automated strategies for estimating the appropriate complexity level.Source: §10.2 · p. 33

• MEDiml-pkg and MEDiml-app connect image processing, standardized feature extraction, feature set reduction, model training, and results analysis through a shared implementation.Source: §11 · p. 33
• In five oncological use cases, the selected complexity level depended on the clinical task, and simpler feature groups sufficed in several evaluated settings.Source: §11 · p. 33
• Code generation makes configured pipelines inspectable and extendable as executable Jupyter Notebooks, while mediml.app provides software, tutorials, and support.Statistical comparisons, performance heatmaps, and the Explanation Tree support examining model performance and contributing features.Source: §11 · p. 33

• Martin Vallières acknowledges support from Unité de soutien au système de santé apprenant (SSA) Québec.Source: Funding information · p. 33