Exploring Patient Data Requirements in Training Effective AI Models for MRI-Based Breast Cancer Classification

2025-02-22

Deep-Breath@MICCAI

AI Summary

This study addresses the problem of determining the minimal clinically feasible dataset size required to train MRI-based AI models for breast cancer classification. Moving beyond conventional image-count–based evaluations, we propose a patient-level data requirements analysis framework and introduce the novel metric “effective patient count.” Using a multicenter MRI dataset, our methodology integrates few-shot learning, cross-site robustness evaluation, and uncertainty-driven data importance ranking to quantitatively assess how dataset size, lesion diversity, and annotation quality impact model generalizability. Results demonstrate that only 80–120 high-quality, expert-annotated patients suffice for models to achieve >92% AUC on external multi-institutional validation—substantially lowering the data acquisition barrier for clinical deployment. Our core contribution is the establishment of a reproducible, patient-centric data efficiency evaluation paradigm, providing empirically grounded guidance for data curation and resource planning in medical imaging AI.

Technology Category

• Machine Learning: Calibration & Uncertainty Quantification
• Computer Vision: Multi-modal Vision
• Humans and AI: Other Foundations of Human Computation & AI

Application Category

• Economics, Online Markets and Human Computation: Data quality aspects of human-annotated datasets
• Search and Retrieval-Augmented AI: Web evaluation methodologies and metrics
• User Modeling, Personalization and Recommendation: Metrics for user behavior and evaluating success

Problem

• Determine data quantity for AI training
• Optimize MRI-based breast cancer detection
• Assess impact of patient count on model performance

Innovation

• Utilizes foundation models effectively
• Trains AI with minimal MRI data
• Enhances performance with simple ensembles