ProsRegNet

ProsRegNet registers histopathology-derived cancer labels onto pre-operative magnetic resonance imaging (MRI) scans for prostate cancer using a deep learning–based image registration framework.

Key Features:

• Deep learning registration: Uses deep neural networks to perform image registration between histopathology images and pre-operative MRIs.
• Image preprocessing: Includes initial preprocessing steps to prepare MRI and histopathology images for registration.
• Affine and deformable transformation estimation: Estimates both affine (rigid) and deformable transformations using deep neural networks for accurate alignment.
• Label mapping: Maps ground truth cancer labels from post-surgical histopathology images onto pre-operative MRI using the estimated transformations.
• Training dataset: Developed and trained on MR and histopathology images from 99 patients (Cohort 1).
• Evaluation cohorts: Evaluated on 53 patients from three cohorts, including 12 patients from Cohort 1 and 41 patients from two public datasets.
• Performance: Demonstrates more accurate registration than state-of-the-art algorithms and operates at least 20× faster.

Scientific Applications:

• Prostate MRI–histopathology integration: Provides precise spatial correspondence between histopathology and MRI for prostate cancer.
• Ground truth label transfer: Enables transfer of histopathology-derived cancer labels to pre-operative MRI for validation and analysis.
• Assessment of radiologic interpretation: Supplies objective reference labels to quantify inter-observer variability and radiologist performance on prostate MRI.
• Tumor detection and characterization: Supports improved detection and localization of clinically significant prostate cancer on MRI by providing histopathology-aligned labels.

Methodology:

Computational steps include image preprocessing, estimation of affine and deformable transformations via deep neural networks, mapping of cancer labels from histopathology to MRI using the estimated transformations, training on a 99-patient cohort (Cohort 1), and evaluation on 53 patients from three cohorts (12 from Cohort 1 and 41 from two public datasets).