LightAutoML: Automated Radiomics Optimization

• LightAutoML is an automated machine learning framework that streamlines radiomics pipelines by integrating image preprocessing, feature extraction, and classification.
• It employs modular optimization strategies like random search and Bayesian methods to simultaneously select algorithms and tune hyperparameters.
• LightAutoML combines traditional handcrafted feature engineering with deep discovery radiomics, enhancing reproducibility and easing clinical application.

Radiomics-specific AutoML frameworks constitute a class of automated machine learning solutions systematically designed and engineered to support radiomics pipelines. Unlike general-purpose AutoML platforms, these frameworks address the radiomics domain’s unique requirements: high-dimensional handcrafted and/or custom image features, robust feature selection, reproducible and harmonized modeling, segmentation integration, and accessibility for clinical and research users often operating with limited programming expertise. Historical radiomics workflows relied on manual selection and parameterization of preprocessing, feature extraction, and classification methods, a process both tedious and prone to irreproducibility. AutoML for radiomics seeks to automate this process by optimizing or even evolving entire pipelines across image modalities, anatomical sites, and endpoints, dramatically reducing investigator workload and promoting method standardization and objective benchmarking.

1. Fundamental Principles of Radiomics-Specific AutoML

Radiomics-specific AutoML frameworks encapsulate both traditional feature-based radiomics methodologies and more recent "discovery radiomics" approaches using deep neural networks. The canonical pipeline (e.g., "WORC" (Starmans et al., 2021)) comprises modular components for image/ROI preprocessing, feature extraction (intensity, shape, texture, orientation descriptors), feature filtering or selection, sample resampling, and classification. A distinguishing principle is end-to-end automation via combined algorithm selection and hyperparameter optimization (CASH): model choice and parameter tuning occur simultaneously, guided by objective (cross-validated) performance metrics such as area under ROC curve (AUC) or weighted F1-score.

Discovery radiomics frameworks, such as the StochasticNet Radiomic Sequencer (Shafiee et al., 2015) and Evolutionary Deep Radiomic Sequencer Discovery (Shafiee et al., 2017), integrate automated feature engineering by learning radiomic representations directly from raw imaging data, embedding feature construction within convolutional architectures whose topology or parameterization is itself optimized or evolved. These mechanisms obviate manual feature design and selection, shifting the optimization frontier to architectural hyperparameters (layer count, filter size, connectivity probability) and training criteria.

2. Workflow Architectures and Optimization Strategies

Framework workflows are organized into directed acyclic computation graphs or sequential modules, with each node or module exposing algorithmic options and tunable parameters. For example, "DARWIN" (Chang et al., 2020) employs a GUI-driven computational graph where users chain data loaders, preprocessing steps, feature selection nodes, and model evaluators in drag-and-drop fashion, optionally leaving hyperparameters blank to trigger AutoML search.

In "WORC" (Starmans et al., 2021), the pipeline search space spans:

• Preprocessing: normalization, group-wise feature dropping, imputation.
• Feature selection: RELIEF, LASSO, SelectFromModel, PCA, univariate tests.
• Resampling: RandomUnder/OverSampling, SMOTE variants, ADASYN.
• Classification: SVM, random forest, logistic regression, ensemble methods.

Optimization strategies include random search, Bayesian optimization (e.g., SMAC), grid search, and evolutionary algorithms. The optimization aims to minimize a loss function $\mathcal{L}$ (e.g., weighted F1) via $k$-fold cross-validation:

$$\lambda_C^* = \arg\min_{\lambda_C\in\Delta_C} \frac{1}{k} \sum_{i=1}^k \mathcal{L}(\text{train}=D^{(i)}_{\text{train}}(\lambda_C), \text{valid}=D^{(i)}_{\text{valid}}(\lambda_C))$$

This formulation allows selection among algorithms and their internal hyperparameters, with optional modules controlled by Bernoulli “activator” hyperparameters.

Discovery radiomics frameworks leverage stochastic and evolutionary topological search. The StochasticNet sequencer (Shafiee et al., 2015) samples random graph filter connectivities ($p=0.5$ in experiments), implicitly performing model averaging and capacity control. EDRS (Shafiee et al., 2017) encodes network weights as probabilistic DNA, imposes environmental compaction constraints, and synthesizes successive network generations to organically evolve toward efficient and discriminative representations.

AutoML search in agentic frameworks such as "mAIstro" (Tzanis et al., 30 Apr 2025) utilizes agent teams (feature extraction, classification, segmentation), with optimization strategies including grid, random, Bayesian, and evolutionary algorithms, and model evaluation coordinated via orchestrator agents.

3. Feature Engineering: Handcrafted and Learned Representations

Traditional radiomics AutoML centers on extraction and selection of handcrafted features. The typical engine (e.g., PyRadiomics in DARWIN (Chang et al., 2020), mAIstro (Tzanis et al., 30 Apr 2025)) supports:

• First-order statistics (mean, variance, skewness, kurtosis, energy, entropy).
• Shape descriptors (volume, surface area, sphericity, compactness).
• Texture matrices: GLCM, GLRLM, GLSZM, GLDM, NGTDM.
• Filter-based features (wavelet transforms, Laplacian-of-Gaussian, Gabor, local binary patterns).

The feature engineering process includes feature selection (variance thresholding, recursive feature elimination, statistical tests) and automated stability analysis such as SULOV (Simplatab (Lozano-Montoya et al., 13 Jan 2026)), though most AutoML frameworks lack formal integration of test-retest reproducibility or harmonization modules.

Discovery radiomics paradigms (StochasticNet (Shafiee et al., 2015), EDRS (Shafiee et al., 2017)) replace handcrafted features with abstract deep features discovered through data-driven training. Here, convolutional filters with stochastic or evolved sparse connectivity encode higher-order spatial relationships, yielding radiomic sequences in fixed-length feature vectors optimized for discrimination. Performance metrics in these frameworks consistently outperform state-of-the-art handcrafted approaches (e.g., SNRS: 91.07% sensitivity, 75.98% specificity, 84.49% accuracy; EDRS: 93.42%, 82.39%, 88.78%, respectively).

4. Pipeline Extensibility: Segmentation, Survival Analysis, and Federated Workflows

Radiomics-specific AutoML frameworks exhibit varied support for upstream and downstream pipeline integration. Segmentation modules are present in mAIstro (Tzanis et al., 30 Apr 2025) (nnU-Net, TotalSegmentator) and DARWIN (Chang et al., 2020) (seeded region-growing, cross-series manipulation), with outputs feeding directly to radiomics extractor agents. Survival analysis remains a notable gap: only AutoPrognosis (Lozano-Montoya et al., 13 Jan 2026) includes time-to-event modeling, but is computationally infeasible for high-dimensional radiomics on standard hardware.

Multi-center validation, federated learning, and harmonization/batch-effect correction are currently lacking in accessible frameworks. Recommendations include embedding automated feature-reproducibility analysis (intraclass correlation coefficients), IBSI-compliant audits, ComBat-style harmonization, and transparent, user-accessible meta-parameterization in both GUI and programmable API interfaces (Lozano-Montoya et al., 13 Jan 2026).

5. Benchmarking, Performance Metrics, and Usability

Recent peer-reviewed benchmarking (Lozano-Montoya et al., 13 Jan 2026, Starmans et al., 2021) assessed radiomics-specific AutoML tools across diverse datasets:

• Simplatab (Lozano-Montoya et al., 13 Jan 2026): AUC 81.81% ± 4.4 (statistically superior to general-purpose AutoML), mean runtime ~1 hour.
• WORC (Starmans et al., 2021): Random search and Top$_{100}$ ensemble strategy; robust performance across 12 clinical applications (AUC 0.45–0.87 depending on modality and endpoint).
• Discovery radiomics (SNRS, EDRS): consistently exceed 84% accuracy, outperforming classical methods and belief decision trees (Shafiee et al., 2015, Shafiee et al., 2017).

Usability and accessibility vary. Most frameworks (AutoRadiomics, Auto-ML for Radiomics, WORC) require advanced scripting, limiting their practical deployment. Simplatab and DARWIN provide no-code/GUI interfaces, lowering ML expertise requirements and presenting advanced interpretability options (SHAP, bias plots). mAIstro (Tzanis et al., 30 Apr 2025) leverages agentic orchestration, natural-language prompts, and modular tool integration for accessible, programmable radiomics modeling.

Framework	Usability	Maintenance Status	Interpretability
Simplatab	No-code GUI	Active	SHAP/bias visualizations
WORC	Advanced scripting	Maintenance	None
DARWIN	Drag-and-drop GUI	Active	Metric/ROC/trace plots
mAIstro	NL interface	Active, open-source	Logs, model inspection
Auto-ML for	Advanced scripting	Obsolete	Basic

6. Limitations, Gaps, and Future Directions

Radiomics-specific AutoML frameworks confront significant limitations:

• Survival models lack efficient integration for high-dimensional imaging data (Lozano-Montoya et al., 13 Jan 2026).
• Feature reproducibility, IBSI compliance, and harmonization steps are rarely automated.
• End-to-end integration (image preprocessing, segmentation QC, harmonization, modeling) remains incomplete in accessible frameworks (Lozano-Montoya et al., 13 Jan 2026).
• Federated learning and domain-adaptation/institutional batch correction are underrepresented (Lozano-Montoya et al., 13 Jan 2026).

Plausible implications suggest future work will exploit agentic orchestration for multi-stage workflows (e.g., mAIstro), integrate advanced optimizers (evolutionary, Bayesian, RL controllers), expand support for segmentation, plug-in scripting, GPU acceleration, and regulatory logging. Extensibility and auditability are increasingly prioritized, with open-source toolkits (WORC, mAIstro) and standardized benchmark datasets fostering transparent comparison and reproducibility (Starmans et al., 2021, Tzanis et al., 30 Apr 2025, Lozano-Montoya et al., 13 Jan 2026).

7. Summary and Outlook

Radiomics-specific AutoML frameworks are systematically reconfiguring the landscape of quantitative imaging analysis. By formalizing the entire method-selection and parameterization process as a structured optimization problem over algorithmic modules and hyperparameters (including deep architectural search), these frameworks deliver robust, reproducible, and performant models across diverse imaging domains and clinical endpoints. The ongoing evolution from handcrafted, manually tuned pipelines to agentic, orchestrated, end-to-end AutoML systems is evident in both research prototypes (mAIstro (Tzanis et al., 30 Apr 2025), DARWIN (Chang et al., 2020)) and robust benchmarking tools (Simplatab (Lozano-Montoya et al., 13 Jan 2026), WORC (Starmans et al., 2021)). However, crucial developmental gaps remain—most notably in survival modeling, reproducibility, harmonization, and federated validation—requiring continued algorithmic and systems innovation in radiomics AutoML research.