Learn To Be Uncertain: Leveraging Uncertain Labels In Chest X-rays With Bayesian Neural Networks

Hao-Yu Yang, Junling Yang, Yue Pan, Kunlin Cao, Qi Song, Feng Gao, Youbing Yin; Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2019, pp. 5-8

Abstract


Communication of uncertainty is important for both radiology reports and deep neural networks (DNNs). For radiologists, conveying diagnostic uncertainty in the written report is a challenging and yet inevitable task. On the other hand, while deep learning models have shown compelling potentials in disease classification and lesion detection, applications of DNNs in the medical domain should provide a quantitative measurement of prediction confidence for risk management purposes. In this paper, we investigate the relationship between uncertainty in diagnostic chest x-ray radiology reports and uncertainty estimation of corresponding DNN models using Bayesian approaches. Two sampling methods, Bernoulli and Gaussian dropout have been tested. Our results show that the incorporation of uncertainty labels during model training results in higher predictive variance for uncertain cases at test time. The uncertain cases are inherently difficult to diagnose for human readers, which often needs a further psychical examination to confirm. Returning uncertain predictions on these cases will prevent the DNN model from making over-confident mistakes.

Related Material


[pdf]
[bibtex]
@InProceedings{Yang_2019_CVPR_Workshops,
author = {Yang, Hao-Yu and Yang, Junling and Pan, Yue and Cao, Kunlin and Song, Qi and Gao, Feng and Yin, Youbing},
title = {Learn To Be Uncertain: Leveraging Uncertain Labels In Chest X-rays With Bayesian Neural Networks},
booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops},
month = {June},
year = {2019}
}