AI-biopsy

AI-biopsy integrates MR imaging and histopathological labels with deep learning to classify prostate lesions and predict cancer risk.

Key Features:

• Pathology-Radiology Fusion: Fuses MR images with histopathological labels, including Gleason Score/Grade Group, to enhance diagnostic signal.
• Deep Learning Models: Employs deep learning models trained on MR images from 400 patients (228 in-house, 172 external) to classify benign versus malignant lesions and high- versus low-risk tumors.
• Performance Metrics: Reports areas under the ROC curve of 0.89 for cancer versus benign classification and 0.78 for high- versus low-risk classification, and provides negative predictive value, positive predictive value, specificity, sensitivity, accuracy, and Cohen's kappa.
• Class Activation Mapping: Applies class activation mapping to highlight MR image regions that contribute to model predictions.

Scientific Applications:

• Non-invasive diagnosis: Enables prostate cancer assessment from MR images labeled with histopathology to support non-invasive diagnostic evaluation.
• Risk stratification: Distinguishes high-risk from low-risk tumors to inform clinical decision-making and treatment planning.
• Biopsy reduction potential: Provides data-driven cancer risk assessments from MR images that can potentially reduce unnecessary biopsies.

Methodology:

Uses MRI datasets from 400 patients (228 in-house, 172 external); MR images were reviewed by radiologists and labeled with biopsy results including Gleason Score/Grade Group under pathologist supervision; deep learning models were trained to classify benign/malignant and high/low-risk categories and evaluated using AUC and other performance metrics, with class activation mapping used for visualization.