Abstract:The importance of quantifying uncertainty in deep networks has become paramount for reliable real-world applications. In this paper, we propose a method to improve uncertainty estimation in medical Image-to-Image (I2I) translation. Our model integrates aleatoric uncertainty and employs Uncertainty-Aware Regularization (UAR) inspired by simple priors to refine uncertainty estimates and enhance reconstruction quality. We show that by leveraging simple priors on parameters, our approach captures more robust uncertainty maps, effectively refining them to indicate precisely where the network encounters difficulties, while being less affected by noise. Our experiments demonstrate that UAR not only improves translation performance, but also provides better uncertainty estimations, particularly in the presence of noise and artifacts. We validate our approach using two medical imaging datasets, showcasing its effectiveness in maintaining high confidence in familiar regions while accurately identifying areas of uncertainty in novel/ambiguous scenarios.

What problem does this paper attempt to address?

### What problem does this paper attempt to solve? This paper aims to solve the problem of **uncertainty quantification in medical image - to - image (I2I) translation tasks**. Specifically, the author proposes a new method to improve uncertainty estimation in medical I2I translation and enhance the reconstruction quality by introducing **Uncertainty - Aware Regularization (UAR)**. #### Main problem background 1. **The necessity of uncertainty**: - In deep neural networks, quantifying uncertainty is crucial for reliable real - world applications. Especially in critical fields such as medical, military, criminal justice, and autonomous driving, the performance of a model is not only about high precision but also needs to be able to identify situations where the model is not confident to ensure safety and reliability. 2. **Challenges in medical I2I translation**: - I2I translation refers to converting an image in one domain to an image in another domain while maintaining semantic consistency and information integrity. For example, in endoscopy, narrow - band imaging (NBI) can enhance the visualization of abnormal areas, but if some areas are not captured using NBI, subsequent diagnosis will be limited. - Deep - learning methods usually generate inherent uncertainty when dealing with I2I translation, especially in blurry or unseen scenarios. This uncertainty can increase rapidly with changes in the dataset or minor changes in the acquisition modality. 3. **Limitations of existing methods**: - Existing multi - stage architectures require sequential uncertainty estimation between different models, which increases complexity and computational cost. - Most existing uncertainty quantification methods cannot distinguish between familiar regions and new or significantly larger sources of uncertainty. #### Main contributions of the paper 1. **Uncertainty - Aware Regularization (UAR)**: - A simple and model - independent UAR method is proposed. By introducing a lightweight regularization term in the optimization process, it directly improves both the uncertainty and the quality of image translation in a single model. - UAR not only improves translation performance but also can more accurately estimate data - driven uncertainty under common noise interference. 2. **New paired dataset**: - A new paired dataset is introduced for RGB - to - FICE (Flexible Spectral Imaging Color Enhancement) I2I translation in capsule endoscopy. #### Method overview - **Probabilistic I2I translation**: By modeling the residual distribution (such as heteroscedastic Gaussian distribution or generalized normal distribution), it allows for estimating the uncertainty in predictions. - **Uncertainty - Aware Regularization**: By using total variation (TV) to penalize the differences between adjacent pixels, it suppresses false uncertainty caused by noise, thus obtaining more accurate uncertainty estimates. #### Experimental results - **Quantitative evaluation**: The effectiveness of UAR is verified on multiple metrics (such as PSNR, SSIM, LPIPS, RRMSE), especially in the presence of different types of noise and artifacts, UAR shows better performance. - **Qualitative evaluation**: Visualization results show that the uncertainty maps generated by UAR are clearer and have more consistent structures, and can better reflect the uncertain regions in the image. In conclusion, by introducing the UAR method, this paper solves the key problems of uncertainty estimation and reconstruction quality in medical I2I translation tasks, and improves the reliability and robustness of the model in practical applications.

Uncertainty-Aware Regularization for Image-to-Image Translation

Uncertainty-Guided Progressive GANs for Medical Image Translation

Uncertainty Estimation in Contrast-Enhanced MR Image Translation with Multi-Axis Fusion

Quantifying and leveraging predictive uncertainty for medical image assessment

PET Denoising and Uncertainty Estimation Based on NVAE Model

On the Quantification of Image Reconstruction Uncertainty without Training Data

Uncertainty Penalized Weighted Least Squares Framework For Pet Reconstruction Under Uncertain System Models

Quantifying Registration Uncertainty With Sparse Bayesian Modelling

Uncertainty quantification via localized gradients for deep learning-based medical image assessments

Improving Image-Based Precision Medicine with Uncertainty-Aware Causal Models

Propagation and Attribution of Uncertainty in Medical Imaging Pipelines

UncertaINR: Uncertainty Quantification of End-to-End Implicit Neural Representations for Computed Tomography

Domain Transfer Through Image-to-Image Translation for Uncertainty-Aware Prostate Cancer Classification

A Simple Framework to Quantify Different Types of Uncertainty in Deep Neural Networks for Image Classification

Improving Uncertainty-Error Correspondence in Deep Bayesian Medical Image Segmentation

Heteroscedastic Uncertainty Estimation Framework for Unsupervised Registration

Calibrating Ensembles for Scalable Uncertainty Quantification in Deep Learning-based Medical Segmentation

Uncertainty Quantification via Neural Posterior Principal Components

Just rotate it! Uncertainty estimation in closed-source models via multiple queries

NeU-NBV: Next Best View Planning Using Uncertainty Estimation in Image-Based Neural Rendering

Approaching Neural Network Uncertainty Realism