Disentangled Uncertainty and Out of Distribution Detection in Medical Generative Models

TL;DR

This paper investigates disentangled uncertainty quantification in medical image translation using CycleGAN, comparing methods like Ensembles and Dropout, and demonstrates epistemic uncertainty's role in out-of-distribution detection for safer AI deployment.

Contribution

It introduces a comprehensive comparison of uncertainty methods in medical image translation and highlights epistemic uncertainty's effectiveness in detecting out-of-distribution data.

Findings

Ensembles and Dropout outperform other methods in uncertainty estimation.

Epistemic uncertainty effectively detects out-of-distribution inputs.

Disentangled uncertainty improves reliability in medical image translation.

Abstract

Trusting the predictions of deep learning models in safety critical settings such as the medical domain is still not a viable option. Distentangled uncertainty quantification in the field of medical imaging has received little attention. In this paper, we study disentangled uncertainties in image to image translation tasks in the medical domain. We compare multiple uncertainty quantification methods, namely Ensembles, Flipout, Dropout, and DropConnect, while using CycleGAN to convert T1-weighted brain MRI scans to T2-weighted brain MRI scans. We further evaluate uncertainty behavior in the presence of out of distribution data (Brain CT and RGB Face Images), showing that epistemic uncertainty can be used to detect out of distribution inputs, which should increase reliability of model outputs.