Leveraging uncertainty information from deep neural networks for disease detection

By Christian Leibig, Vaneeda Allken, Murat Seckin Ayhan, Philipp Berens, Siegfried Wahl. 2016

553 citations recorded by the source corpus.

Deep learning (DL) has revolutionized the field of computer vision and image processing. In medical imaging, algorithmic solutions based on DL have been shown to achieve high performance on tasks that previously required medical experts. However, DL-based solutions for disease detection have been proposed without methods to quantify and control their uncertainty in a decision. In contrast, a physician knows whether she is uncertain about a case and will consult more experienced colleagues if needed. Here we evaluate drop-out based Bayesian uncertainty measures for DL in diagnosing diabetic retinopathy (DR) from fundus images and show that it captures uncertainty better than straightforward alternatives. Furthermore, we show that uncertainty informed decision referral can improve diagnostic performance. Experiments across different networks, tasks and datasets show robust generalization. Depending on network capacity and task/dataset difficulty, we surpass 85% sensitivity and 80% specificity as recommended by the NHS when referring 0−20% of the most uncertain decisions for further inspection. We analyse causes of uncertainty by relating intuitions from 2D visualizations to the high-dimensional image space. While uncertainty is sensitive to clinically relevant cases, sensitivity to unfamiliar data samples is task dependent, but can be rendered more robust.

Subfields: Artificial Intelligence, Radiology, Nuclear Medicine and Imaging, Medical & Health AI

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Artificial Intelligence

In Radiology, Nuclear Medicine and Imaging

In Medical & Health AI

Citations in the corpus

An overview of deep learning in medical imaging focusing on MRI (2018, background): “In addition to producing valuable measures that function as uncertainty measures [65,355,356], these techniques can also lessen deep neural networks susceptibility to adversarial attacks [354,357].”
Opportunities and obstacles for deep learning in biology and medicine (2017, background): “[45] estimated the uncertainty of deep networks for diabetic retinopathy diagnosis by linking dropout networks with approximate Bayesian inference.”
Towards neural networks that provably know when they don't know (2019, methodology): “However, several cases of high conﬁdence predictions far away from the training data have been reported for neural networks, e.g. fooling images (Nguyen et al., 2015), for out-of-distribution (OOD) images (Hendrycks & Gimpel, 2017a) or in medical diagnosis (Leibig et al., 2017).”
Why ReLU Networks Yield High-Confidence Predictions Far Away From the Training Data and How to Mitigate the Problem (2018, methodology): “Moreover, note that the quite useful models for conﬁdence calibration on the input distribution like [10, 12, 18] cannot be used for out-of-distribution detection as it has been observed in [21].”
Radial Bayesian Neural Networks: Beyond Discrete Support In Large-Scale Bayesian Deep Learning (2019, methodology): “We perform classification on a dataset of ‘fundus’ images taken of the back of retinas in order to diagnose diabetic retinopathy [Kaggle, 2015] building on Leibig et al. [2017] and Filos et al. [2019]. Diabetic retinopathy is graded in five stages, where 0 is healthy and 4 is”
What Do Compressed Deep Neural Networks Forget (2019, methodology): “Our PIE methodology could conceivably be explored as a mechanism to surface a tractable subset of atypical examples for further human inspection (Leibig et al., 2017; Zhang, 1992), to choose not to classify certain examples when the model is uncertain (Bartlett & Wegkamp, 2008; Cortes et al.”
A Systematic Comparison of Bayesian Deep Learning Robustness in Diabetic Retinopathy Tasks (2019, methodology): “[30], we augment training dataset using affine transformations, including random zooming (by up to ±10%), random translations (independent shifts by up to ±25 pixels) and random rotations (by up to ±π).”
A Survey of Deep Active Learning (2020, methodology): “In addition, in practice, researchers often also use a powerful and low-cost Monte-Carlo dropout (MC-dropout) [156] stochastic regularization technique to obtain posterior samples, consequently attaining good performance on real-world datasets [80, 94].”
Exploring Uncertainty Measures in Deep Networks for Multiple Sclerosis Lesion Detection and Segmentation (2018, methodology): “In [10], the authors perform an evaluation of the MC sample variance for image-based diabetic retinopathy diagnosis.”
Selective Brain Damage: Measuring the Disparate Impact of Model Pruning (2019, background): “PIE could be used to surface atypical examples for further human inspection [48], choose not to classify certain examples when the model is uncertain [2, 10, 9, 11], or to aid interpretability as a case based reasoning tool to explain model behavior [41, 23, 4, 33].”
CODEs: Chamfer Out-of-Distribution Examples against Overconfidence Issue (2021, methodology): “Besides, it has been validated that models for confidence calibration on the input distribution cannot be used for out-of-distribution [29].”
Confidence Calibration and Predictive Uncertainty Estimation for Deep Medical Image Segmentation (2019, methodology): “nearities, which provide an approximation to a probabilistic Gaussian process. Implementing MC dropout is straightforward and has been applied in several application domains including medical imaging [18]. In a similar Bayesian approach, Teye et al. [19] showed that training NNs with BN [7] can be u”
Deep learning in bioinformatics: introduction, application, and perspective in big data era (2019, background): “diction results.The uncertainty scaling is even more important in healthcare, which can help us evaluate the reliability of machine learning-based disease diagnosis and an automatic clinical decision [79].The probability score from the direct deep learning Softmax output is usually not in the right ”
On the Interpretability of Artificial Intelligence in Radiology: Challenges and Opportunities. (2020, methodology): “Uncertainty estimation has also been used to assist in the referral of wrongly classified medical images for disease detection (30).”
Distinction Maximization Loss: Fast, Scalable, Turnkey, and Native Neural Networks Out-of-Distribution Detection simply by Replacing the SoftMax Loss (2019, background): “awa, 2019) or even ensemble methods (Vyas et al., 2018; Lakshminarayanan et al., 2017). Finally, there are proposals based on uncertainty or conﬁdence estimation/calibration (Kendall & Gal, 2017; Leibig et al., 2017; Malinin & Gales, 2018; Kuleshov et al., 2018; Subramanya et al., 2017). Des”
A Review of Uncertainty Quantification in Deep Learning: Techniques, Applications and Challenges (2020, unknown): “[45] 2017 CNN diabetic retinopathy √”
An Exploratory Analysis of Multi-Class Uncertainty Approximation in Bayesian Convolutional Neural Networks (2018, methodology): “In Part 2 we explore the empirical properties of these ad hoc methods, with particular focus on the work presented in [Leibig et al., 2017].”
Understanding Measures of Uncertainty for Adversarial Example Detection (2018, methodology): “Some works in the literature estimate the epistemic uncertainty of a dropout model using the estimated variance of the MC samples, rather than the mutual information (Leibig et al., 2017; Feinman et al., 2017; Carlini & Wagner, 2017a).”