Research map / Probabilistic and scientific modeling / Uncertainty and Gaussian processes

PersonalizedUS: Interpretable Breast Cancer Risk Assessment with Local Coverage Uncertainty Quantification

AAAI 2025 · Probabilistic and scientific modeling › Uncertainty and Gaussian processes · approach: Regression uncertainty quantification · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Builds PersonalizedUS, an interpretable logistic-regression breast ultrasound malignancy risk system that adds Locart local-coverage conformal prediction sets, benchmarked on a 1936-lesion multicenter dataset.

Problem

Assessing malignancy of breast lesions found on ultrasound is hard; BI-RADS scoring causes unnecessary biopsies, and existing ML models lack interpretability and personalized uncertainty quantification.

Approach

Applies Locart (feature-space partition via a decision tree fit to classification residuals, then split conformal calibration per leaf) to binary malignancy classification, yielding interpretable logistic regression risk estimates with locally valid prediction sets.

Key results

• Logistic regression achieved sensitivity 0.900 and specificity 0.888 on the calibration set, with AUROC 0.958 and log-loss 0.268.
• Locart prediction sets at alpha=0.1 gave average set size near 1 for most leaves, with empirical coverage of 89-100% per leaf; leaves 11 and 12 were the hard subgroups.
• On BI-RADS 4a/4b test lesions, the threshold t=0.223 would have requested 64 of 196 biopsies (>65% reduction) with no missed cancers.

Measured effects reported in the paper

Surprising findings

• More complex learners (neural networks, XGBoost, random forest) showed only marginal gains over logistic regression, so interpretability was prioritized.

Limitations

• Dataset is from a small set of Brazilian centers and the test set is prospective-only with 364 lesions.
• Conditional coverage is only approximate, with observed empirical coverage fluctuating up to 10% per leaf due to ~100 points per leaf.
• Code and data release deferred until after review.

Experimental setup

Datasets: Breast Lesion Dataset (1936 biopsied breast lesions, multicenter Brazilian study)

Benchmarks: Breast Lesion Dataset

Metrics: sensitivity specificity PPV NPV AUROC AUPRC log-loss average set size empirical coverage truth-only set proportion

Baselines: decision tree Naive Bayes k-nearest neighbors XGBoost neural network AdaBoost random forest support vector machine

Builds on: Locart conformal prediction BI-RADS

Closest papers by meaning

• PH-Net: Semi-Supervised Breast Lesion Segmentation via Patch-wise Hardness (CVPR 2024 · similarity 0.82)
• The Boundaries of Fair AI in Medical Image Prognosis: A Causal Perspective (NeurIPS 2025 · similarity 0.82)
• TumorChain: Interleaved Multimodal Chain-of-Thought Reasoning for Traceable Clinical Tumor Analysis (ICLR 2026 · similarity 0.81)
• Samba: Severity-aware Recurrent Modeling for Cross-domain Medical Image Grading (NeurIPS 2024 · similarity 0.81)
• Enhancing Treatment Effect Estimation via Active Learning: A Counterfactual Covering Perspective (ICML 2025 · similarity 0.81)
• "Why Is There a Tumor?": Tell Me the Reason, Show Me the Evidence (ICML 2025 · similarity 0.81)