VIDS-Seg: Towards Reliable Uncertainty Quantification in Pediatric Cardiac Ultrasound Segmentation

Clinical machine learning models often suffer from silent failures when deployed on populations that differ from their training data, such as applying adult-trained cardiac segmentation models to pediatric patients. The authors investigate whether uncertainty quantification (UQ) can serve as a reliable safety layer to detect these failures at inference time without requiring expensive retraining on labeled pediatric data.

The authors introduce VIDS-Seg, an extension of the Variational Inference under Distribution Shifts (VIDS) framework tailored for dense image segmentation. Instead of applying variational inference to an entire deep U-Net—which is computationally prohibitive—the authors freeze the U-Net backbone and apply amortized variational inference only to a lightweight 1x1 convolutional prediction head. This head is conditioned on a context summary of the training data and the specific test image, allowing the model to generate an adaptive prior that increases uncertainty when it encounters out-of-distribution (OOD) inputs.

VIDS-Seg was evaluated using EchoNet-Dynamic (adults) for training and EchoNet-Pediatric (children) for zero-shot testing. The results demonstrate that VIDS-Seg maintains high segmentation accuracy while providing superior spatial correspondence between its uncertainty maps and actual segmentation errors. This advantage remains robust even when compared against standard baselines that use post-hoc temperature scaling. Furthermore, the improved uncertainty estimates lead to more stable and accurate ejection fraction calculations, particularly for the infant subgroup, which represents the most significant anatomical shift from the adult training data.