Voxel-wise prostate cell density prediction using multiparametric magnetic resonance imaging and machine learning

Area of Science:

• Medical Imaging
• Computational Biology
• Oncology

Background:

• No current methods exist to estimate prostate cell density.
• Prostate cancer research requires accurate cell density quantification.

Purpose of the Study:

• To develop predictive models for estimating prostate cell density.
• Utilize multiparametric magnetic resonance imaging (mpMRI) data.
• Employ machine learning techniques at a voxel level.

Main Methods:

• Collected in vivo mpMRI data from 30 patients.
• Co-registered mpMRI with histology-derived cell density maps.
• Applied multivariate adaptive regression spline (MARS), polynomial regression (PR), and generalised additive model (GAM) algorithms.
• Optimized models using leave-one-out cross-validation and evaluated with root mean square error (RMSE).

Main Results:

• Successfully trained and tested predictive models for voxel-wise prostate cell density.
• The best performing model, a generalised additive model (GAM), achieved an RMSE of 1.06 × 10³ cells/mm².
• The optimal model demonstrated a relative deviation of 13.3%.

Conclusions:

• Prostate cell density can be quantitatively estimated non-invasively from mpMRI.
• High-quality co-registered data at a voxel level is crucial for accurate predictions.
• These estimations can aid in tissue classification, treatment response assessment, and personalized radiotherapy.