Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Reliable Multi-View Learning with Conformal Prediction for Aortic Stenosis Classification in Echocardiography

Ang Nan Gu, Michael Tsang, Hooman Vaseli, Teresa Tsang, Purang Abolmaesumi

TL;DR

This work tackles uncertainty in AS classification from echocardiography, where 2-D views of a 3-D anatomy can omit diagnostic details. It introduces Re-Training for Uncertainty (RT4U), a model-agnostic training approach that uses pseudo-labels derived from training dynamics to mitigate overfitting to noisy inputs, and couples RT4U with Conformal Prediction to produce adaptive prediction sets with guaranteed coverage. Across CIFAR-Q, TMED-2, and a private AS dataset, RT4U improves top-1 accuracy and calibration while yielding smaller, more informative prediction sets, preserving the marginal coverage property. The approach offers a practical, uncertainty-aware pathway for scalable, automated AS screening from standard echo data, bridging data-centric training with principled uncertainty quantification.

Abstract

The fundamental problem with ultrasound-guided diagnosis is that the acquired images are often 2-D cross-sections of a 3-D anatomy, potentially missing important anatomical details. This limitation leads to challenges in ultrasound echocardiography, such as poor visualization of heart valves or foreshortening of ventricles. Clinicians must interpret these images with inherent uncertainty, a nuance absent in machine learning's one-hot labels. We propose Re-Training for Uncertainty (RT4U), a data-centric method to introduce uncertainty to weakly informative inputs in the training set. This simple approach can be incorporated to existing state-of-the-art aortic stenosis classification methods to further improve their accuracy. When combined with conformal prediction techniques, RT4U can yield adaptively sized prediction sets which are guaranteed to contain the ground truth class to a high accuracy. We validate the effectiveness of RT4U on three diverse datasets: a public (TMED-2) and a private AS dataset, along with a CIFAR-10-derived toy dataset. Results show improvement on all the datasets.

Reliable Multi-View Learning with Conformal Prediction for Aortic Stenosis Classification in Echocardiography

TL;DR

This work tackles uncertainty in AS classification from echocardiography, where 2-D views of a 3-D anatomy can omit diagnostic details. It introduces Re-Training for Uncertainty (RT4U), a model-agnostic training approach that uses pseudo-labels derived from training dynamics to mitigate overfitting to noisy inputs, and couples RT4U with Conformal Prediction to produce adaptive prediction sets with guaranteed coverage. Across CIFAR-Q, TMED-2, and a private AS dataset, RT4U improves top-1 accuracy and calibration while yielding smaller, more informative prediction sets, preserving the marginal coverage property. The approach offers a practical, uncertainty-aware pathway for scalable, automated AS screening from standard echo data, bridging data-centric training with principled uncertainty quantification.

Abstract

The fundamental problem with ultrasound-guided diagnosis is that the acquired images are often 2-D cross-sections of a 3-D anatomy, potentially missing important anatomical details. This limitation leads to challenges in ultrasound echocardiography, such as poor visualization of heart valves or foreshortening of ventricles. Clinicians must interpret these images with inherent uncertainty, a nuance absent in machine learning's one-hot labels. We propose Re-Training for Uncertainty (RT4U), a data-centric method to introduce uncertainty to weakly informative inputs in the training set. This simple approach can be incorporated to existing state-of-the-art aortic stenosis classification methods to further improve their accuracy. When combined with conformal prediction techniques, RT4U can yield adaptively sized prediction sets which are guaranteed to contain the ground truth class to a high accuracy. We validate the effectiveness of RT4U on three diverse datasets: a public (TMED-2) and a private AS dataset, along with a CIFAR-10-derived toy dataset. Results show improvement on all the datasets.
Paper Structure (18 sections, 3 equations, 3 figures, 2 tables)

This paper contains 18 sections, 3 equations, 3 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Clinical motivation.
  3. Challenges in interpreting echo exam data.
  4. Conformal prediction.
  5. Contributions.
  6. Background and Related Works
  7. Methodology
  8. Training dynamics with input-dependent noise
  9. RT4U algorithm
  10. Experiments
  11. Simulated noise on CIFAR-10 dataset
  12. Aortic stenosis classification using image and video data
  13. Metrics
  14. Decision aggregation
  15. Results and discussion
  16. ...and 3 more sections

Figures (3)

  • Figure 1: Conformal prediction returns a set $\{.\}$ of predictions. The set size correlates with the estimated uncertainty. Our RT4U method (green) enables better prediction and more adaptive set sizes compared to prior art (blue). Left: Inference on CIFAR-10 quadrants, orange box indicates the quadrant seen by the network. Right: Inference on echo cross-sections from the TMED-2 dataset. Ground truth is embedded in each echo image.
  • Figure 2: Demonstration of the RT4U algorithm in three steps. 1: When fitting on one-hot labels, the performance over different epochs vary depending on the clarity of the image (blue line). 2: Using the average of the past predictions, we form a pseudo-label (orange dotted line). 3: Fitting on the new pseudo-label improves stability and the model's uncertainty-awareness (green line).
  • Figure 3: Left: Calibration performance of R(2+1)D variants with RT4U and temperature scaling, evaluated at the instance-level on the AS Private dataset. Right: Histograms of network confidence for the predicted class. Compared to the baseline (blue), RT4U (green) has a lower expected calibration error, and the confidence distribution is closer to those from models with temperature scaling.