Optimizing radiomics for prostate cancer diagnosis: feature selection strategies, machine learning classifiers, and MRI sequences

Abstract

Objectives: Radiomics-based analyses encompass multiple steps, leading to ambiguity regarding the optimal approaches for enhancing model performance. This study compares the effect of several feature selection methods, machine learning (ML) classifiers, and sources of radiomic features, on models' performance for the diagnosis of clinically significant prostate cancer (csPCa) from bi-parametric MRI.

Methods: Two multi-centric datasets, with 465 and 204 patients each, were used to extract 1246 radiomic features per patient and MRI sequence. Ten feature selection methods, such as Boruta, mRMRe, ReliefF, recursive feature elimination (RFE), random forest (RF) variable importance, L1-lasso, etc., four ML classifiers, namely SVM, RF, LASSO, and boosted generalized linear model (GLM), and three sets of radiomics features, derived from T2w images, ADC maps, and their combination, were used to develop predictive models of csPCa. Their performance was evaluated in a nested cross-validation and externally, using seven performance metrics.

Results: In total, 480 models were developed. In nested cross-validation, the best model combined Boruta with Boosted GLM (AUC = 0.71, F1 = 0.76). In external validation, the best model combined L1-lasso with boosted GLM (AUC = 0.71, F1 = 0.47). Overall, Boruta, RFE, L1-lasso, and RF variable importance were the top-performing feature selection methods, while the choice of ML classifier didn't significantly affect the results. The ADC-derived features showed the highest discriminatory power with T2w-derived features being less informative, while their combination did not lead to improved performance.

Conclusion: The choice of feature selection method and the source of radiomic features have a profound effect on the models' performance for csPCa diagnosis.

Critical relevance statement: This work may guide future radiomic research, paving the way for the development of more effective and reliable radiomic models; not only for advancing prostate cancer diagnostic strategies, but also for informing broader applications of radiomics in different medical contexts.

Key points: Radiomics is a growing field that can still be optimized. Feature selection method impacts radiomics models' performance more than ML algorithms. Best feature selection methods: RFE, LASSO, RF, and Boruta. ADC-derived radiomic features yield more robust models compared to T2w-derived radiomic features.

Keywords: MRI; Machine learning; Prostate cancer; Radiomics.

Fig. 1

The workflow of the study in four steps

Fig. 2

Boxplots of the AUC and F1 score for all the combinations of feature selection methods and ML classifiers in setting 1. The average performance (red points) is provided on the right side of the boxes

Fig. 3

Boxplots of the AUC and F1 score for all the combinations of feature selection methods and ML classifiers in setting 2. The average performance (red points) is provided on the right side of the boxes

Fig. 4

Models exhibiting statistically significant differences in ROC AUC in setting 2 and the frequency at which they outperformed other models in Delong’s test

Fig. 5

Variation of AUC explained by feature selection method, ML classifiers, and MRI sequence, and their interactions (A) for all settings, and (B) for settings 1 and 2, separately

Fig. 6

Comparative analysis of feature selection methods for ADC radiomic features across settings and folds. (A) Barplot depicting the number of selected features for each method, and (B) Radar plot illustrating the selection frequency for each feature aggregated across the ten feature selection methods