Dataset Information

Confidence Calibration and Predictive Uncertainty Estimation for Deep Medical Image Segmentation.

ABSTRACT: Fully convolutional neural networks (FCNs), and in particular U-Nets, have achieved state-of-the-art results in semantic segmentation for numerous medical imaging applications. Moreover, batch normalization and Dice loss have been used successfully to stabilize and accelerate training. However, these networks are poorly calibrated i.e. they tend to produce overconfident predictions for both correct and erroneous classifications, making them unreliable and hard to interpret. In this paper, we study predictive uncertainty estimation in FCNs for medical image segmentation. We make the following contributions: 1) We systematically compare cross-entropy loss with Dice loss in terms of segmentation quality and uncertainty estimation of FCNs; 2) We propose model ensembling for confidence calibration of the FCNs trained with batch normalization and Dice loss; 3) We assess the ability of calibrated FCNs to predict segmentation quality of structures and detect out-of-distribution test examples. We conduct extensive experiments across three medical image segmentation applications of the brain, the heart, and the prostate to evaluate our contributions. The results of this study offer considerable insight into the predictive uncertainty estimation and out-of-distribution detection in medical image segmentation and provide practical recipes for confidence calibration. Moreover, we consistently demonstrate that model ensembling improves confidence calibration.

SUBMITTER: Mehrtash A

PROVIDER: S-EPMC7704933 | biostudies-literature |

REPOSITORIES: biostudies-literature

ACCESS DATA

Similar Datasets

Project description:Problem. Image biomarker analysis, also known as radiomics, is a tool for tissue characterization and treatment prognosis that relies on routinely acquired clinical images and delineations. Due to the uncertainty in image acquisition, processing, and segmentation (delineation) protocols, radiomics often lack reproducibility. Radiomics harmonization techniques have been proposed as a solution to reduce these sources of uncertainty and/or their influence on the prognostic model performance. A relevant question is how to estimate the protocol-induced uncertainty of a specific image biomarker, what the effect is on the model performance, and how to optimize the model given the uncertainty. Methods. Two non-small cell lung cancer (NSCLC) cohorts, composed of 421 and 240 patients, respectively, were used for training and testing. Per patient, a Monte Carlo algorithm was used to generate three hundred synthetic contours with a surface dice tolerance measure of less than 1.18 mm with respect to the original GTV. These contours were subsequently used to derive 104 radiomic features, which were ranked on their relative sensitivity to contour perturbation, expressed in the parameter η. The top four (low η) and the bottom four (high η) features were selected for two models based on the Cox proportional hazards model. To investigate the influence of segmentation uncertainty on the prognostic model, we trained and tested the setup in 5000 augmented realizations (using a Monte Carlo sampling method); the log-rank test was used to assess the stratification performance and stability of segmentation uncertainty. Results. Although both low and high η setup showed significant testing set log-rank p-values (p = 0.01) in the original GTV delineations (without segmentation uncertainty introduced), in the model with high uncertainty, to effect ratio, only around 30% of the augmented realizations resulted in model performance with p < 0.05 in the test set. In contrast, the low η setup performed with a log-rank p < 0.05 in 90% of the augmented realizations. Moreover, the high η setup classification was uncertain in its predictions for 50% of the subjects in the testing set (for 80% agreement rate), whereas the low η setup was uncertain only in 10% of the cases. Discussion. Estimating image biomarker model performance based only on the original GTV segmentation, without considering segmentation, uncertainty may be deceiving. The model might result in a significant stratification performance, but can be unstable for delineation variations, which are inherent to manual segmentation. Simulating segmentation uncertainty using the method described allows for more stable image biomarker estimation, selection, and model development. The segmentation uncertainty estimation method described here is universal and can be extended to estimate other protocol uncertainties (such as image acquisition and pre-processing).

Project description:Recent advances in deep learning for medical image segmentation demonstrate expert-level accuracy. However, application of these models in clinically realistic environments can result in poor generalization and decreased accuracy, mainly due to the domain shift across different hospitals, scanner vendors, imaging protocols, and patient populations etc. Common transfer learning and domain adaptation techniques are proposed to address this bottleneck. However, these solutions require data (and annotations) from the target domain to retrain the model, and is therefore restrictive in practice for widespread model deployment. Ideally, we wish to have a trained (locked) model that can work uniformly well across unseen domains without further training. In this paper, we propose a deep stacked transformation approach for domain generalization. Specifically, a series of n stacked transformations are applied to each image during network training. The underlying assumption is that the "expected" domain shift for a specific medical imaging modality could be simulated by applying extensive data augmentation on a single source domain, and consequently, a deep model trained on the augmented "big" data (BigAug) could generalize well on unseen domains. We exploit four surprisingly effective, but previously understudied, image-based characteristics for data augmentation to overcome the domain generalization problem. We train and evaluate the BigAug model (with n=9 transformations) on three different 3D segmentation tasks (prostate gland, left atrial, left ventricle) covering two medical imaging modalities (MRI and ultrasound) involving eight publicly available challenge datasets. The results show that when training on relatively small dataset (n = 10~32 volumes, depending on the size of the available datasets) from a single source domain: (i) BigAug models degrade an average of 11%(Dice score change) from source to unseen domain, substantially better than conventional augmentation (degrading 39%) and CycleGAN-based domain adaptation method (degrading 25%), (ii) BigAug is better than "shallower" stacked transforms (i.e. those with fewer transforms) on unseen domains and demonstrates modest improvement to conventional augmentation on the source domain, (iii) after training with BigAug on one source domain, performance on an unseen domain is similar to training a model from scratch on that domain when using the same number of training samples. When training on large datasets (n = 465 volumes) with BigAug, (iv) application to unseen domains reaches the performance of state-of-the-art fully supervised models that are trained and tested on their source domains. These findings establish a strong benchmark for the study of domain generalization in medical imaging, and can be generalized to the design of highly robust deep segmentation models for clinical deployment.

Dataset Information

Confidence Calibration and Predictive Uncertainty Estimation for Deep Medical Image Segmentation.

Similar Datasets

OmicsDI is part of the ELIXIR infrastructure

Tweets