Source-linked AI summary
A Review of Uncertainty Estimation and its Application in Medical Imaging
Ke Zou, Zhihao Chen, Xuedong Yuan, Xiaojing Shen, Meng Wang, Huazhu Fu
TL;DR
Medical-imaging AI can be unreliable when clinical data differ from training data, creating a need for confidence information alongside predictions. This review synthesizes uncertainty types, estimation methods, and medical-imaging applications, then identifies challenges including missing uncertainty ground truth, computational complexity, accuracy–reliability trade-offs, and limited clinical evaluation.
Problem
Medical-imaging AI deployment is limited by insufficient reliability and trustworthiness of predictions in real clinical settings.
Method
The paper reviews aleatoric and epistemic uncertainty, estimation methods, uncertainty-aware medical-imaging models, applications, challenges, and future directions.
Results
The review organizes uncertainty estimation across medical-imaging tasks and identifies deterministic, Bayesian, ensemble, and test-time augmentation approaches alongside application-specific evidence.
Takeaways & Limitations
Uncertainty estimation can provide confidence information for identifying problematic inputs and supporting clinician involvement in critical cases.
Takeaways & Limitations
Clinical evaluation remains limited, making it difficult to compare methods and determine which are most effective and efficient across tasks.
Abstract
from arXiv · showhide
The use of AI systems in healthcare for the early screening of diseases is of great clinical importance. Deep learning has shown great promise in medical imaging, but the reliability and trustworthiness of AI systems limit their deployment in real clinical scenes, where patient safety is at stake. Uncertainty estimation plays a pivotal role in producing a confidence evaluation along with the prediction of the deep model. This is particularly important in medical imaging, where the uncertainty in the model's predictions can be used to identify areas of concern or to provide additional information to the clinician. In this paper, we review the various types of uncertainty in deep learning, including aleatoric uncertainty and epistemic uncertainty. We further discuss how they can be estimated in medical imaging. More importantly, we review recent advances in deep learning models that incorporate uncertainty estimation in medical imaging. Finally, we discuss the challenges and future directions in uncertainty estimation in deep learning for medical imaging. We hope this review will ignite further interest in the community and provide researchers with an up-to-date reference regarding applications of uncertainty estimation models in medical imaging.
I. INTRODUCTION
Deep learning supports medical imaging tasks but can become unreliable when clinical inputs differ from training data. Uncertainty estimation adds confidence information to help identify unreliable predictions and guide clinician involvement.
- Deep learning is used for medical imaging tasks including tumor segmentation, disease diagnosis, and treatment planning, but introduces risks such as bias, errors, and limited transparency.
- Closed-world assumptions may fail in real clinical scenes, causing performance drops and unreliable predictions on low-quality or out-of-distribution inputs.
- Uncertainty estimation predicts confidence alongside the model output, enabling users to quantify reliability and identify when the system may be performing poorly.
- In interventional applications, high-uncertainty results should be referred to experts so critical decisions are not based solely on uncertain predictions.
- Non-interventional applications use uncertainty to flag out-of-distribution and abnormal cases for further analysis, human intervention, or additional training data.
A. Types of Uncertainty
Deep-learning uncertainty represents limited confidence in a prediction and is commonly divided into aleatoric uncertainty from data noise and epistemic uncertainty from limited model or distribution knowledge.
- Uncertainty measures a model’s lack of confidence or ambiguity about the correct output for a given input.
- Aleatoric uncertainty arises from inherent data noise, such as measurement error or ambiguous annotation, and cannot be reduced by collecting more data.
- Aleatoric uncertainty can be estimated by training a model to output a distribution over possible predictions rather than a single point estimate.
- Epistemic uncertainty reflects limited knowledge about the model or data distribution and can be estimated with Bayesian neural networks or model ensembles.
- Epistemic uncertainty can theoretically decrease with more complex models, additional data, or regularization techniques.
B. Methods of Uncertainty Estimation
The review groups uncertainty-estimation approaches into deterministic, Bayesian, ensemble, and test-time augmentation methods. These methods differ in computational cost, modeling assumptions, and how prediction variation is measured.
- Deterministic methods estimate uncertainty with a single value from one forward pass, offering simplicity and computational efficiency.
- Bayesian neural networks represent model parameters as random variables and estimate uncertainty from distributions of possible outputs.
- Dropout can approximate Bayesian inference by sampling different model architectures during training or prediction.
- Ensemble methods use multiple models and treat the variance of their individual predictions as an uncertainty measure.
- Test-time data augmentation perturbs test inputs and estimates uncertainty from changes in the model’s predictions across those inputs.
A. Classification
Medical image classification has achieved strong predictive performance, but reliability analysis remains important for clinical deployment. Reviewed uncertainty-aware approaches span ophthalmology, histopathology, skin lesions, MRI, and chest radiographs, while scalability and labeling remain challenges.
- Classification methods can achieve clinician-comparable results in some areas, yet limited reliability analysis continues to restrict clinical deployment.
- Reviewed applications cover ophthalmology, histopathology, skin disease, MRI, and chest-radiograph images.
- Dropout-based Bayesian uncertainty measures captured diagnostic uncertainty in diabetic-retinopathy fundus-image classification without requiring additional uncertainty labels.
- A trusted multi-scale whole-slide-image framework estimated uncertainty at different magnifications and integrated evidence across magnifications for classification.
- Uncertainty-aware studies also applied dropout, distribution-overlap measures, Bayesian convolutional networks, and variational inference across histopathology, skin-lesion, and MRI classification.
- Chest-radiograph datasets contain many uncertain samples, motivating label smoothing and further research on converting uncertain labels into definite labels.
- Most prior studies rely on stochastic, inefficient MC-Dropout approaches, while many remain task-specific and lack end-to-end scalability.
B. Segmentation
Medical image segmentation uncertainty estimation spans Bayesian and non-Bayesian approaches, with methods trading computational cost, sampling requirements, and segmentation performance for robustness and OOD detection.
- Segmentation uncertainty captures ambiguity near tissue boundaries and uncertainty about a region’s semantic category.
- Bayesian methods learn distributions over network weights, while non-Bayesian approaches include MC dropout, ensembles, and deterministic models.
- MC dropout and ensemble methods require multiple forward passes or additional training, increasing computational burden during uncertainty estimation.
- Deterministic methods produce uncertainty in a single test-time forward pass, improving efficiency but potentially sacrificing segmentation performance.
- Future work targets more robust and calibrated uncertainty, using uncertainty to improve segmentation, and incorporating it into training.
C. Other tasks
Uncertainty estimation extends beyond classification and segmentation to registration, reconstruction, denoising, super-resolution, counting, detection, and tumor growth prediction. Registration and reconstruction receive the most attention because uncertainty can inform operative risk, while several other tasks remain less studied.
- Uncertainty estimation has been applied to registration, reconstruction, denoising, super-resolution, counting, detection, and tumor growth prediction.
- For image registration, uncertainty helps surgeons assess surgical risk from the reliability of registered images, including transformation and label uncertainty.
- MRI reconstruction methods estimate uncertainty while restoring high-fidelity images from partial measurements, including through adaptive measurement selection and iterative refinement.
- Super-resolution methods combine heteroscedastic noise modeling and approximate Bayesian inference to quantify predictive uncertainty.
- Registration and reconstruction have received the most attention, whereas uncertainty estimation is less studied for tasks such as denoising and counting.
IV. DISCUSSION AND CONCLUSION
The review identifies missing uncertainty ground truth, computational complexity, accuracy–reliability trade-offs, and limited clinical evaluation as major challenges. Addressing these constraints is necessary to assess and deploy uncertainty estimation effectively in medical imaging.
- Many medical applications lack ground truth for uncertainty, making estimation accuracy and method evaluation difficult.
- Computational complexity can hinder scaling uncertainty estimation to real-time clinical systems.
- A model may have high accuracy but low reliability, so both metrics should be considered and their trade-off depends on the use case.
- Empirical evaluation remains limited, especially in real clinical scenarios, complicating comparisons across methods and tasks.