A data-driven approach to prostate cancer detection from dynamic contrast enhanced MRI

Area of Science:

• Radiology
• Medical Imaging
• Oncology

Background:

• Dynamic contrast-enhanced magnetic resonance imaging (DCE-MRI) is crucial for prostate cancer diagnosis and staging.
• Current DCE-MRI analysis relies on pharmacokinetic models that often oversimplify perfusion and require challenging arterial input function estimation.
• Limitations in existing models necessitate more robust methods for accurate prostate cancer characterization.

Purpose of the Study:

• To develop and validate a data-driven approach for prostate cancer detection using DCE-MRI data.
• To evaluate the efficacy of model-free empirical parameters and principal component analysis (PCA) features against traditional pharmacokinetic parameters.
• To improve multiparametric MRI protocols for enhanced prostate cancer detection.

Main Methods:

• A data-driven analysis of DCE-MRI time series was performed, bypassing traditional pharmacokinetic modeling.
• Model-free empirical parameters and PCA of normalized T1-weighted intensities were extracted as features.
• Least absolute shrinkage and selection operator (LASSO) regression identified optimal principal components, and a support vector machine classifier was used for detection.

Main Results:

• The proposed data-driven approach achieved an area under the ROC curve of 0.86.
• This performance surpasses the 0.78 AUC obtained with traditional pharmacokinetic parameters.
• Validation was conducted on 449 tissue regions from 16 patients, with results correlated to histological evaluations.

Conclusions:

• The novel data-driven method using LASSO-isolated PCA parameters shows superior performance in prostate cancer detection compared to conventional pharmacokinetic parameters.
• This approach offers a promising alternative for analyzing DCE-MRI data, potentially enhancing multiparametric MRI protocols.
• The findings suggest a significant advancement in non-invasive prostate cancer diagnosis and staging.