Radiomics Reproducibility in Prostate Cancer Diagnosis Based on PROSTATEx.

Area of Science:

• Radiology
• Medical Imaging
• Machine Learning

Background:

• Prostate cancer (PCa) diagnosis often relies on invasive methods.
• Noninvasive diagnostic tools are crucial for early detection and management.
• Radiomics and machine learning offer potential for improved PCa diagnosis.

Purpose of the Study:

• To extract radiomics features from prostate MRI.
• To evaluate the reproducibility of these features.
• To develop machine learning models using reproducible features for noninvasive PCa diagnosis.

Main Methods:

• Prostate MRI data from 82 subjects (41 PCa, 41 controls) were analyzed.
• Radiomics features were extracted from T2-weighted imaging (T2WI) and apparent diffusion coefficient (ADC) maps.
• Reproducibility was assessed using the intraclass correlation coefficient (ICC ≥ 0.75), and features were selected, normalized, and reduced.

Main Results:

• Reproducible radiomics features significantly contributed to model performance.
• Machine learning models (SVM, NN, LR) achieved 80-84% accuracy and 0.85 AUC.
• Principal component analysis provided more consistent results than nonlinear dimensionality reduction methods.

Conclusions:

• Combining reproducible MRI radiomics features with ML provides a robust noninvasive method for PCa diagnosis.
• Emphasis on feature reproducibility improves model performance and reliability.
• This approach supports potential clinical translation for prostate cancer diagnostics.