UMORSS: Uncertainty-aware Multimodal Ovarian Risk Scoring System

This repository contains the code implementation for UMORSS (Uncertainty-aware Multimodal Ovarian Risk Scoring System), a novel multimodal AI framework for automated ovarian cancer risk assessment that integrates uncertainty quantification. UMORSS combines deep learning-based ultrasound image analysis with clinical biomarkers to provide reliable prediction of malignancy risk while accounting for model uncertainty.

Environment Setup

Install requirements:

pip install -r requirements.txt

Project Structure

• models/ - Model architecture implementations
  • van.py and van2.py - Vision Attention Network (VAN) model
• single_test.py - Testing script for single case prediction
• ai-assistance.py - Helper functions for combining doctor and AI predictions
• LASSO.csv - Feature coefficients from LASSO regression

Usage

1. Single case testing:

python single_test.py

This script demonstrates prediction on a single test image with:

• Phase 1 initial risk assessment using VAN model
• Phase 2 detailed analysis combining ultrasound imaging and clinical features
• Uncertainty estimation

2. AI-Doctor combined prediction:

python ai-assistance.py

Provides functions to combine O-RADS scores from:

• Doctor's assessment
• AI model predictions
• Uncertainty measurements

Model Details

The system uses a two-phase approach:

1. Initial screening using VAN for binary risk classification
2. Detailed analysis combining:
   • Deep learning features from ultrasound image
   • Clinical biomarkers and patient data
   • LASSO regression for feature selection