🤖 machine learning

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

This paper introduces VIDS-Seg, a framework that leverages amortized variational inference to provide reliable uncertainty quantification for pediatric cardiac ultrasound segmentation, enabling the detection of silent failures in zero-shot scenarios where models trained on adult data are applied to children without requiring retraining.

Original authors: Paul Fischer, Ece Ozkan

Published 2026-08-12

📖 7 min read🧠 Deep dive

CC BY 4.0

Original authors: Paul Fischer, Ece Ozkan

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are teaching a robot to recognize cats. You show it a thousand pictures of fluffy, adult cats sitting on sofas. The robot gets really good at this. But then, you hand it a picture of a tiny, hairless kitten with giant ears. Because the robot was only trained on adult cats, it might squint at the kitten, confidently say, "That's definitely a cat," and draw a perfect box around it. The problem? The robot is wrong, but it doesn't know it's wrong. It's confident, but it's failing silently. This is a huge problem in the world of medical AI. Doctors use these robots to help look at X-rays and ultrasounds, but if the robot was trained mostly on adults, it might get confused when looking at a baby's heart. It might make a mistake and not tell the doctor, which could lead to bad medical decisions. Scientists call this "Out-of-Distribution" (OOD) failure. They also want the robot to have "Uncertainty Quantification," which is just a fancy way of saying the robot should know when it is guessing and say, "Hey, I'm not sure about this one!"

This paper introduces a new method called VIDS-Seg to solve exactly that problem. The researchers wanted to see if they could make a medical AI model that knows when it is looking at something it wasn't trained on—like a baby's heart when it was trained on adults—and actually raise a red flag instead of confidently making a mistake. They tested this by training a model on adult heart ultrasound videos and then seeing how it performed on children of different ages. They found that the model struggled the most with infants, who have very different heart shapes and sizes compared to adults. While other popular AI methods kept confidently making mistakes on these babies, VIDS-Seg was able to spot the trouble spots. It didn't just guess; it highlighted exactly where it was confused, creating a "worry map" that matched the errors. This means that in the real world, a doctor could look at this map, see the robot is unsure, and double-check the baby's heart manually, preventing a silent failure.

The Story of the Robot That Learned to Say "I Don't Know"

In the world of medical imaging, we have super-smart computer programs that can look at pictures of the heart and draw lines around the chambers. It's like having a digital artist that never gets tired. But there's a catch: these artists are usually trained on pictures of grown-ups. When you show them a picture of a baby, things get tricky. Babies aren't just tiny adults; their hearts are shaped differently, they beat faster, and the ultrasound pictures look different.

The researchers in this paper asked a simple question: What happens when our digital artist tries to draw a baby's heart?

They took a model trained on over 10,000 adult heart scans (from a dataset called EchoNet-Dynamic) and asked it to draw the hearts of children from a different dataset (EchoNet-Pediatric). They split the children into groups: teenagers, school-age kids, toddlers, and infants.

The Surprise Discovery:
The robot was pretty good at drawing the hearts of teenagers and older kids. But when it got to the infants (babies under one year old), it started to stumble. The accuracy dropped, and the errors got bigger. The researchers realized that infants are so different from adults that they are essentially "Out-of-Distribution" (OOD) for the robot. The robot was trying to force a baby's heart into an adult's shape, and it was failing silently.

The Problem with "Confident" Mistakes

Here is the scary part: most AI models are terrible at admitting when they are wrong. If you ask a standard AI to draw a baby's heart, it might draw a line that looks okay, but it's actually in the wrong place. And the worst part? The AI will tell you, "I am 99% sure I got this right!"

The researchers tested three different types of AI to see which one could admit when it was confused:

The Ensemble: A team of 10 different robots voting on the answer.
PHiSeg: A robot designed to guess where the lines are fuzzy.
VIDS-Seg: The new robot built by the authors, based on a framework called VIDS.

They measured how well each robot's "worry level" (uncertainty) matched its actual mistakes. They used a score called NCC (Normalized Cross-Correlation) to see if the robot was worried about the right spots.

The Results:

PHiSeg was the worst. It didn't know where to worry. Its "worry map" was tiny and only looked at the edges, missing the big mistakes in the middle.
The Ensemble was better, but it still got a lot of the baby hearts wrong without raising enough alarms.
VIDS-Seg was the clear winner. It produced a "worry map" that perfectly highlighted the areas where it was making mistakes. When the robot was confused, the map got bright and noisy. When it was sure, the map was calm.

Even when the researchers tried to "calibrate" the other robots (like adjusting the volume on a radio to make the sound clearer), VIDS-Seg still won. The other robots were just confused in the wrong places; VIDS-Seg was confused in the right places.

Why This Matters for Real Doctors

You might think, "So the robot is better at drawing lines. Big deal." But the real test was what happened next. Doctors use these drawings to calculate something called the Ejection Fraction (EF). This is a percentage that tells you how well the heart is pumping blood. If the EF is too low, the baby might have a serious heart problem.

The researchers checked how accurate the EF numbers were for the babies:

The PHiSeg robot gave very messy, inconsistent EF numbers.
The Ensemble was better but still had a lot of variation.
VIDS-Seg gave the most stable and accurate EF numbers for the babies.

More importantly, when they asked the robots to flag babies with heart problems (EF under 50%), VIDS-Seg was the best at catching them. It had a success rate (AUROC) of 0.94, compared to 0.90 for the Ensemble and 0.81 for PHiSeg.

The Takeaway

The paper shows that we don't always need to retrain AI models with thousands of new baby pictures to make them safe. Instead, we can give them a "safety layer" that knows when they are stepping out of their comfort zone.

VIDS-Seg works by creating a special "adaptive prior." Think of it like a robot that keeps a mental map of what it has seen before. When it sees a baby's heart, it realizes, "Wait, this doesn't look like the adult hearts I studied. I need to be extra careful." It doesn't just guess; it highlights the uncertainty.

The authors suggest that this could be a practical safety tool. If a doctor sees a bright, noisy "worry map" from VIDS-Seg, they know to double-check the baby's heart manually. It's like having a robot that says, "I think I can do this, but I'm not 100% sure, so please look over my shoulder."

This isn't a magic wand that fixes all AI problems, and the authors admit that more testing with real doctors is needed. But it suggests a path forward: making AI models that are honest about their limitations, especially when they are looking at the most vulnerable patients.

Technical Summary: VIDS-Seg for Reliable Uncertainty Quantification in Pediatric Cardiac Ultrasound

Problem Statement
The reliable clinical deployment of machine learning models is hindered by their tendency to fail silently when applied to subgroups underrepresented in training data. A critical example is pediatric care, where models trained predominantly on adult cohorts often exhibit systematic under-performance on children due to significant differences in anatomy, physiology, and image acquisition. While retraining on labeled pediatric data is ideal, it is often infeasible due to data scarcity and the heterogeneity of the pediatric population (e.g., rapid anatomical changes from infancy to adolescence). Consequently, there is a pressing need for methods that can detect these "silent failures" at inference time without requiring additional labeled data or model retraining. Standard uncertainty quantification (UQ) methods often fail in this regard, as they are typically calibrated on in-distribution data and do not inherently increase uncertainty for out-of-distribution (OOD) inputs, leading to confident but incorrect predictions.

Methodology: VIDS-Seg
The authors introduce VIDS-Seg, an adaptation of the VIDS (Variational Inference under Distribution Shifts) framework to dense image segmentation. The core innovation lies in making the adaptive prior of VIDS tractable for high-dimensional segmentation networks.

Architecture Decomposition: Recognizing that applying variational inference to a full U-Net (with $10^7$ – $10^8$ parameters) is computationally intractable, the authors decompose the network into two components:
1. Embedding Network ( $g_\xi$ ): A frozen U-Net backbone (encoder, bottleneck, and decoder) pre-trained on the segmentation task. It extracts rich spatial features but remains fixed during the uncertainty training phase.
2. Prediction Head ( $f_\theta$ ): A lightweight $1\times1$ convolutional layer that maps feature maps to class logits. Variational inference is applied only to the parameters of this head, reducing the dimensionality of the variational posterior to a manageable size.
Adaptive Prior and Training: The method employs an adaptive prior $p(\theta | x_{1:N}, x^*)$ conditioned on both the training context and the test input.
- Inference Network: A neural network $h_\gamma$ takes a summary of the training context (mean and standard deviation of embedded training images) and the embedding of the test image to output the parameters ( $\mu, \sigma$ ) of the variational distribution $q_\phi(\theta)$ .
- Training Strategy: Since shifted test covariates are unavailable during training, the method constructs $L$ synthetic environments by drawing bootstrap subsamples from the training data. The inference network is optimized using a cross-environment Evidence Lower Bound (ELBO) objective with a variance penalty to encourage robustness against distribution shifts.
- Energy Function: The energy function is adapted for segmentation by averaging log-probabilities over spatial locations rather than summing them, ensuring scale independence across different image resolutions.
Inference: At test time, the model generates a context summary from the training set and an embedding for the test image. The inference network produces the variational posterior parameters, from which $S$ samples of the prediction head are drawn. The final segmentation is the mean prediction, and the uncertainty is estimated via the pixel-wise entropy map of these samples.

Key Contributions

VIDS-Seg Framework: Extends the VIDS framework to dense segmentation by applying amortized variational inference over a lightweight prediction head, preserving OOD-awareness while remaining computationally feasible.
Empirical OOD Identification: Systematically evaluates uncertainty across age-stratified pediatric subgroups, identifying infants as the primary OOD population where adult-trained models consistently under-perform.
Downstream Clinical Impact: Demonstrates that improved uncertainty localization translates to more accurate and stable ejection fraction (EF) estimates and better detection of cardiac malfunction in the infant subgroup.
Positioning vs. Calibration: Shows that VIDS-Seg's advantages are not merely due to better-scaled confidence (which post-hoc calibration can fix) but reflect a genuine improvement in the spatial correspondence between uncertainty and error.

Results
The method was evaluated on left ventricular segmentation using EchoNet-Dynamic (adults, training) and EchoNet-Pediatric (children, zero-shot evaluation).

Segmentation Performance: All models (Deep Ensemble, PHiSeg, VIDS-Seg) showed high accuracy (DSC > 0.90) for older children but a significant performance drop (DSC $\approx$ 0.84–0.85) for infants, confirming infants as the OOD group.
Uncertainty Quality: VIDS-Seg achieved the highest Normalized Cross-Correlation (NCC) between the predicted uncertainty (entropy) and the actual segmentation error.
- On the infant (OOD) group, VIDS-Seg achieved an NCC of 0.62 (after temperature scaling), significantly outperforming the Deep Ensemble (0.54) and PHiSeg (0.33).
- Crucially, this superiority persisted even after applying temperature scaling to all baselines, indicating that VIDS-Seg provides better spatial localization of uncertainty, not just better magnitude calibration.
Qualitative Analysis: Entropy maps from VIDS-Seg were tightly concentrated around actual error regions in infant cases, whereas baselines produced uncertainty that was either too narrow (PHiSeg) or missed critical error-prone regions (Ensemble).
Downstream Impact: For the infant subgroup, VIDS-Seg yielded the lowest Mean Absolute Error (MAE) in Ejection Fraction estimation and the highest AUROC (0.94) for detecting cardiac malfunction (EF < 50%), outperforming the Deep Ensemble (AUROC 0.90) and PHiSeg (AUROC 0.81).

Significance and Claims
The paper claims that OOD-aware uncertainty quantification can serve as a practical "safety layer" for deployed segmentation models. By explicitly modeling the dependence of uncertainty on distributional novelty, VIDS-Seg enables the detection of silent failures in underrepresented subgroups (specifically infants) without retraining or additional labeled data.

The authors emphasize that this approach addresses a fundamental limitation of standard UQ methods: the inability to distinguish between in-distribution and out-of-distribution inputs. The results suggest that improving the quality and localization of uncertainty estimates has direct, measurable consequences for clinical decision-making, specifically in the reliability of derived biomarkers like ejection fraction. The work positions VIDS-Seg not as a replacement for calibration, but as a superior base estimator that can be combined with post-hoc methods to enhance clinical safety.

The Story of the Robot That Learned to Say "I Don't Know"

The Problem with "Confident" Mistakes

Why This Matters for Real Doctors

The Takeaway

Technical Summary: VIDS-Seg for Reliable Uncertainty Quantification in Pediatric Cardiac Ultrasound

More like this