Addressing Deep Learning Model Calibration Using Evidential Neural Networks And Uncertainty-Aware Training

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Machine Learning

Background:

• Deep learning (DL) models achieve high accuracy in medical image classification.
• However, DL model outputs, often using SoftMax, can be poorly calibrated (over-confident).
• Evidential neural networks (ENNs) and uncertainty-aware training are proposed solutions.

Purpose of the Study:

• To investigate the calibration improvements from uncertainty-aware training and ENNs, individually and combined.
• To evaluate these methods on a simple MNIST task and a complex medical imaging task.

Main Methods:

• Experiments were conducted on MNIST digit classification.
• Phase Contrast Cardiac Magnetic Resonance images were used for a complex artefact detection task.
• The study compared individual and combined applications of ENNs and uncertainty-aware training.

Main Results:

• Model calibration can degrade with increased task complexity requiring higher-capacity models.
• Combining ENNs and uncertainty-aware training improved calibration for both low and high-capacity models in the complex artefact detection task.

Conclusions:

• The combination of ENNs and uncertainty-aware training shows promise for improving DL model calibration in complex medical imaging applications.
• Improved calibration can enhance clinician trust in DL-driven medical diagnoses.