Pseudo-D: Informing Multi-View Uncertainty Estimation with Calibrated Neural Training Dynamics

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

Keywords: Multi-View Learning · Echocardiography · Uncertainty · Training Dynamics

Abstract:

Computer-aided diagnosis systems must make critical decisions
from medical images that are often noisy, ambiguous, or conflicting,
yet today’s models are trained on overly simplistic labels that ignore diagnostic
uncertainty. One-hot labels erase inter-rater variability and force
models to make overconfident predictions, especially when faced with incomplete
or artifact-laden inputs. We address this gap by introducing
a novel framework that brings uncertainty back into the label space.
Our method leverages neural network training dynamics (NNTD) to assess
the inherent difficulty of each training sample. By aggregating and
calibrating model predictions during training, we generate uncertaintyaware
pseudo-labels that reflect the ambiguity encountered during learning.
This label augmentation approach is architecture-agnostic and can
be applied to any supervised learning pipeline to enhance uncertainty
estimation and robustness. We validate our approach on a challenging
echocardiography classification benchmark, demonstrating superior performance
over specialized baselines in calibration, selective classification,
and multi-view fusion

Submission Number: 15