Abstract
Breast ultrasound (US) interpretation is challenged by high false-positive rates and significant inter-reader variability. We present PersonalizedUS, an interpretable AI system that integrates regularized logistic regression with a novel "local" conformal prediction method (Locart). By partitioning patients into distinct subgroups based on lesion characteristics, PersonalizedUS provides personalized risk assessments with rigorous uncertainty guarantees. In a study of 1,936 biopsied lesions, the system achieved a 65% reduction in unnecessary biopsies for suspicious cases while maintaining 100% sensitivity and an AUC of 0.958.
Introduction
The current standard for breast ultrasound interpretation, the BI-RADS scoring system, often leads to excessive invasive procedures. Studies indicate that only 7-8% of US-guided biopsies result in a malignancy diagnosis, causing unnecessary patient anxiety and high healthcare costs. While deep learning models offer high accuracy, their "black box" nature and lack of rigorous uncertainty measures hinder clinical adoption. PersonalizedUS addresses these issues by providing a transparent, statistically reliable framework for risk assessment.
Methods
The PersonalizedUS system employs a two-stage architecture for diagnostic support:
- Risk Estimation: A regularized logistic regression model calculates the probability of malignancy. Unlike complex neural networks, this model provides clear, weighted contributions for each clinical and ultrasound feature (e.g., age, lesion size, margins), ensuring interpretability.
- Local Uncertainty Quantification (Locart): To provide rigorous confidence measures, the system utilizes Locart, a conformal prediction method adapted for binary classification. Locart uses a decision tree to partition the patient population into subgroups based on case difficulty. It then calculates uncertainty quantiles for each subgroup individually, ensuring that a 90% reliability guarantee holds "locally" for specific patient types rather than just as a global average.
The system is deployed as a web application where clinicians input eight key features to receive a risk percentage, a list of top contributing features, and a "prediction set" (e.g., {Benign}, {Malignant}, or {Benign, Malignant}) reflecting the model's confidence.
Results
The model was validated on a multicenter dataset of 1,936 biopsied breast lesions. The logistic regression approach proved highly competitive with more complex models like Random Forests and XGBoost, achieving a sensitivity of 0.90 and specificity of 0.88.
Key performance highlights include:
- Biopsy Reduction: In simulations of BI-RADS 4a and 4b cases, the system achieved a 65% reduction in requested biopsies while maintaining 100% sensitivity (zero missed cancers).
- High AUC: The model achieved an Area Under the ROC Curve (AUC) of 0.958.
- Reliable Confidence: For 80% of test cases, the model was confident enough to provide a "singleton" prediction set while maintaining its 90% local coverage guarantee.
Discussion and Conclusion
PersonalizedUS demonstrates that interpretability and high diagnostic performance are not mutually exclusive. By using a transparent logistic regression model, clinicians can "debug" the AI's reasoning in real-time. Furthermore, the local uncertainty quantification via Locart identifies "hard" cases where the model is less certain, signaling the need for expert review.
Future work will focus on automating feature extraction from raw ultrasound images and validating the system across more diverse global populations. PersonalizedUS represents a significant step toward reliable, interpretable AI in clinical oncology.
Summary updated