Voxel-based supervised machine learning of peripheral zone prostate cancer using noncontrast multiparametric MRI

Area of Science:

• Radiology
• Medical Imaging
• Machine Learning

Background:

• Accurate prostate cancer detection is crucial for effective treatment.
• Noncontrast multiparametric MRI (mp-MRI) offers a promising alternative to contrast-enhanced imaging.
• Machine learning techniques can enhance the diagnostic capabilities of mp-MRI.

Purpose of the Study:

• To develop and evaluate a supervised machine learning model for classifying cancerous and noncancerous voxels using noncontrast mp-MRI.
• To assess the performance of models utilizing T2-weighted imaging (T2WI), diffusion-weighted imaging (DWI), and diffusion tensor imaging (DTI) parameters.
• To generate cancer probability maps for improved image interpretation and treatment management.

Main Methods:

• Extracted 191 radiomic features from mp-MRI data of prostate cancer patients.
• Developed Support Vector Machine (SVM) models combining T2WI, DWI, and DTI parameters.
• Optimized model parameters using a Bayesian approach and validated with a leave-one-patient-out method.
• Evaluated diagnostic performance using Area Under the Receiver Operating Characteristic Curve (AUROC), sensitivity, specificity, and accuracy.

Main Results:

• The combined T2WI + DWI + DTI model demonstrated the highest performance with an average AUROC of 0.93 ± 0.03.
• Achieved average sensitivity, specificity, and accuracy of 0.85 ± 0.05, 0.82 ± 0.07, and 0.83 ± 0.04, respectively.
• T2WI + DTI models showed slightly better performance than T2WI + DWI models.

Conclusions:

• Supervised classification techniques combined with noncontrast mp-MRI (T2WI, DWI, DTI) and Bayesian optimization can accurately differentiate cancerous from noncancerous voxels.
• This approach provides high diagnostic accuracy without the need for contrast agents.
• Cancer probability maps enhance image interpretation, lesion heterogeneity assessment, and treatment planning.