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

Area of Science:

• Medical imaging analysis
• Machine learning in oncology
• Radiomics

Background:

• Radiomics analysis involves multiple steps, creating ambiguity in optimizing model performance.
• Standardized approaches are needed to enhance diagnostic accuracy in clinical settings.

Purpose of the Study:

• To compare the impact of various feature selection methods, machine learning (ML) classifiers, and radiomic feature sources on diagnostic model performance for clinically significant prostate cancer (csPCa).
• To identify optimal strategies for developing robust radiomics models for csPCa detection using bi-parametric MRI.

Main Methods:

• Utilized two multi-centric datasets (465 and 204 patients) with bi-parametric MRI data.
• Extracted 1246 radiomic features from T2w images, ADC maps, and their combination.
• Evaluated ten feature selection methods (e.g., Boruta, RFE, L1-lasso) and four ML classifiers (SVM, RF, LASSO, GLM) using nested cross-validation and external validation.

Main Results:

• The choice of feature selection method and radiomic feature source significantly affected model performance.
• Boruta, RFE, L1-lasso, and Random Forest variable importance were top-performing feature selection methods.
• ADC-derived features demonstrated higher discriminatory power than T2w-derived features; combined features did not improve performance.

Conclusions:

• Feature selection methods and radiomic feature sources are critical for optimizing radiomics models in csPCa diagnosis.
• ADC-derived radiomic features provide more robust models compared to T2w-derived features.
• This research guides future radiomics development for improved prostate cancer diagnostics and broader medical applications.