Uncertainty-Calibrated Explainable AI for Fetal Ultrasound Plane Classification

TL;DR

This work tackles unreliable fetal ultrasound plane classification caused by noise and distribution shift by proposing an uncertainty-calibrated explainable AI framework. The approach unifies uncertainty estimation (e.g., $f_\theta(x) \in \Delta^{K-1}$ with $T$ stochastic passes, ensembles, or evidential outputs), calibration (temperature scaling and conformal prediction), and explainability (Grad-CAM++ and LIME) with uncertainty reporting. It leverages the FETAL_PLANES_DB benchmark to define a clinician-facing reporting protocol that combines accuracy, calibration (e.g., $ECE$ and $Brier$), selective prediction, and explainability metrics, and supports human-in-the-loop decision-making through escalation rules. By integrating these components into a practical workflow and QC framework, the blueprint aims to deliver trustworthy, reproducible fetal-plane classifiers that maintain trustworthy confidence and informative explanations under noisy acquisition conditions.

Abstract

Fetal ultrasound standard-plane classification underpins reliable prenatal biometry and anomaly screening, yet real-world deployment is limited by domain shift, image noise, and poor calibration of predicted probabilities. This paper presents a practical framework for uncertainty-calibrated explainable AI in fetal plane classification. We synthesize uncertainty estimation methods (Monte Carlo dropout, deep ensembles, evidential learning, and conformal prediction) with post-hoc and uncertainty-aware explanations (Grad-CAM variants, LIME-style local surrogates, and uncertainty-weighted multi-resolution activation maps), and we map these components to a clinician-facing workflow. Using FETAL_PLANES_DB as a reference benchmark, we define a reporting protocol that couples accuracy with calibration and selective prediction, including expected calibration error, Brier score, coverage-risk curves, and structured error analysis with explanations. We also discuss integration points for quality control and human-in-the-loop review, where uncertainty flags trigger re-acquisition or expert confirmation. The goal is a reproducible, clinically aligned blueprint for building fetal ultrasound classifiers whose confidence and explanations remain trustworthy under noisy acquisition conditions.

Figures (1)

• Figure 1: A minimal clinical workflow for uncertainty-calibrated explainable plane classification. Calibrated uncertainty supports selective prediction and escalation, while explanations and quality control provide traceability and user feedback.