Prostate cancer is a leading male malignancy, and early diagnosis is critical -- the 10-year survival rate reaches 99% when the disease is caught and treated early. While biopsy remains the definitive diagnosis tool, MRI (magnetic resonance imaging) plays a pivotal role in identifying suspicious areas before tissue is sampled.
Radiomics is an AI-driven approach that extracts hundreds or thousands of quantitative features from medical images. These features capture subtle patterns in tumor texture and structure that are invisible to the naked eye, helping clinicians detect, diagnose, and characterize cancers like prostate cancer without invasive procedures.
A full radiomics pipeline involves multiple steps: image acquisition, region-of-interest delineation, feature extraction, feature selection, and model building. Each step introduces choices, and it has been unclear which combinations of methods lead to the most reliable results -- a critical gap this study aims to fill.
A particular challenge is that radiomics datasets contain huge numbers of features, many of which are redundant or correlated with each other. Without careful feature selection, machine learning models can overfit training data and fail to generalize to new patients -- undermining their clinical usefulness.
The researchers used two independent datasets: ProstateNET (465 patients from 12 European clinical centers) and ProstateX2 (204 patients scanned on Siemens 3-Tesla machines). Clinically significant prostate cancer (csPCa) was defined as a Gleason grade group of 2 or higher, equivalent to a Gleason score of 7 or above -- the threshold where treatment is generally recommended.
From each patient's MRI, the team extracted 1,246 radiomic features per imaging sequence using the open-source PyRadiomics library. Features came from shape, intensity statistics, and texture descriptors (including co-occurrence matrices and wavelet transforms), derived from two MRI sequences: T2-weighted (T2w) images and apparent diffusion coefficient (ADC) maps from diffusion-weighted imaging.
Ten different feature selection methods were evaluated, spanning three categories: filter methods (mRMRe, ReliefF, CorrSF, SES), wrapper methods (Boruta, RFE, genetic algorithm), and embedded methods (L1-lasso, HSIC-lasso, RF variable importance). Four machine learning classifiers were tested: support vector machine (SVM), random forest (RF), LASSO, and boosted generalized linear model (Boosted GLM).
Models were evaluated in two settings: nested cross-validation within ProstateNET, and external validation using ProstateX2 as an independent test set. Performance was measured by the area under the ROC curve (AUC) and F1 score, among other metrics. In total, 480 unique models were built and compared.
Filter methods rank features based on statistical properties like correlation or mutual information, without building a predictive model. They are fast and simple but evaluate features in isolation, missing interactions between features. Examples include mRMRe (which maximizes relevance while minimizing redundancy) and ReliefF (which scores features by comparing nearest neighbors of different classes).
Wrapper methods such as Boruta and recursive feature elimination (RFE) evaluate subsets of features by actually training a model on them. They capture interactions between features and tend to produce more informative subsets, but at greater computational cost. Boruta works by comparing each feature's importance against randomly permuted shadow versions of the same features.
Embedded methods like L1-lasso and random forest variable importance (RF-imp) perform feature selection as part of the model training itself. L1-lasso applies a penalty that drives the coefficients of less informative features to zero, effectively removing them. RF-imp ranks features by how much they reduce prediction error within a random forest.
All methods were applied only to the training data to prevent data leakage -- a critical precaution ensuring that information from test patients never influences which features are selected. Pre-filtering steps removed features with near-zero variance and those with Pearson correlation above 0.85 before the main selection step.
In nested cross-validation, the best model combined Boruta feature selection with Boosted GLM, achieving an AUC of 0.71 and F1 score of 0.76. In external validation, the best model used L1-lasso with Boosted GLM (AUC = 0.71, F1 = 0.47). The drop in F1 score in external validation reflects the harder task of generalizing to a new patient population with a much lower prevalence of clinically significant cancer (25% vs. 74%).
Across all experiments, ADC-derived radiomic features consistently outperformed T2w-derived features. Combining T2w and ADC features (biparametric MRI) did not improve performance compared to ADC alone -- suggesting that the added complexity of T2w features introduces noise rather than useful information.
The top-performing feature selection methods were Boruta, RFE, L1-lasso, and RF variable importance. These consistently produced better models regardless of which classifier was used. By contrast, the choice of machine learning classifier (SVM, RF, LASSO, or Boosted GLM) had minimal effect on performance -- all four performed similarly.
A statistical analysis of what drives AUC variability found that MRI sequence choice explained 28.8% of total variance, feature selection method explained 5.8%, and the choice of classifier explained less than 2%. This clearly establishes MRI sequence and feature selection as the dominant factors -- not the choice of algorithm.
Even though different feature selection methods identified different subsets of features, there was a core group of 16 ADC radiomic features selected across all settings and folds by at least one method. This consistency suggests these features capture genuinely informative biological signals in the tumor's diffusion properties.
The most consistently selected feature was "wavelet.LLL.firstorder_Minimum" -- a feature derived from a low-frequency wavelet transform of the ADC map, capturing the minimum intensity value in that transformed image. It was selected by all methods in external validation and 8 out of 10 methods in each cross-validation fold, making it a candidate biomarker for clinically significant prostate cancer.
The number of features selected varied widely by method: Boruta and RFE tended to select 17-26 features, while SES selected only 4-6. L1-lasso showed the highest variability (3-27 features). Smaller feature sets are generally preferred for clinical translation as they reduce model complexity and improve interpretability.
The finding that feature selection matters more than classifier choice is practically important: researchers spending effort tuning algorithms like SVM vs. random forest may be focusing on the wrong variable. The study confirms that investing in robust feature selection methods -- particularly RFE, Boruta, L1-lasso, and RF variable importance -- is more valuable.
The superiority of ADC features over T2w features aligns with biological logic: ADC maps reflect how freely water molecules diffuse within tissue, a property directly related to cell density and tumor aggressiveness. T2w imaging, while useful for anatomy, appears to add less predictive power when ADC is already included.
The study followed the CLEAR checklist (44/58 items met) and received an "Excellent" quality score (81.7%) on the METRICS radiomics quality scoring system. This methodological rigor is rare in radiomics literature and increases confidence in the study's conclusions.
Limitations include the fact that the pre-filtering step (removing correlated and irrelevant features before the main selection) was not itself evaluated as a variable. Incorporating clinical variables like PSA levels or patient age into the models might also improve performance and warrants future study.
This study provides actionable guidance for radiomics researchers: when building models for prostate cancer diagnosis, prioritize ADC-derived features and use robust feature selection methods like Boruta, RFE, L1-lasso, or RF variable importance. These choices have far more impact on model quality than which classifier is used.
The findings are not limited to prostate cancer. The general principle -- that feature selection method and imaging source dominate model performance -- likely applies to radiomics across many cancer types and imaging modalities. This makes the study relevant as a methodological benchmark for the broader field.
Future work should explore deep learning-based feature selection (such as autoencoders), which may capture non-linear patterns that traditional methods miss. Combining radiomics with clinical variables is another avenue that could push performance beyond the AUC of 0.71 achieved here. Multi-center external validation across new scanner types and geographies is also needed before clinical deployment.