DropConnect is effective in modeling uncertainty of Bayesian deep networks

Uncertainty quantification in multi-class segmentation: Comparison between Bayesian and non-Bayesian approaches in a clinical perspective

BACKGROUND

Automatic segmentation techniques based on Convolutional Neural Networks (CNNs) are widely adopted to automatically identify any structure of interest from a medical image, as they are not time consuming and not subject to high intra- and inter-operator variability. However, the adoption of these approaches in clinical practice is slowed down by some factors, such as the difficulty in providing an accurate quantification of their uncertainty.

PURPOSE

This work aims to evaluate the uncertainty quantification provided by two Bayesian and two non-Bayesian approaches for a multi-class segmentation problem, and to compare the risk propensity among these approaches, considering CT images of patients affected by renal cancer (RC).

METHODS

Four uncertainty quantification approaches were implemented in this work, based on a benchmark CNN currently employed in medical image segmentation: two Bayesian CNNs with different regularizations (Dropout and DropConnect), named BDR and BDC, an ensemble method (Ens) and a test-time augmentation (TTA) method. They were compared in terms of segmentation accuracy, using the Dice score, uncertainty quantification, using the ratio of correct-certain pixels (RCC) and incorrect-uncertain pixels (RIU), and with respect to inter-observer variability in manual segmentation. They were trained with the Kidney and Kidney Tumor Segmentation Challenge launched in 2021 (Kits21), for which multi-class segmentations of kidney, RC, and cyst on 300 CT volumes are available. Moreover, they were tested considering this and other two public renal CT datasets.

RESULTS

Accuracy results achieved large differences across the structures of interest for all approaches, with an average Dice score of 0.92, 0.58, and 0.21 for kidney, tumor, and cyst, respectively. In terms of uncertainties, TTA provided the highest uncertainty, followed by Ens and BDC, whereas BDR provided the lowest, and minimized the number of incorrect certain pixels worse than the other approaches. Again, large differences were seen across the three structures in terms of RCC and RIU. These metrics were associated with different risk propensity, as BDR was the most risk-taking approach, able to provide higher accuracy in its prediction, but failing to assign uncertainty on incorrect segmentation in every case. The other three approaches were more conservative, providing large uncertainty regions, with the drawback of giving alert also on correct areas. Finally, the analysis of the inter-observer segmentation variability showed a significant variation among the four approaches on the external dataset, with BDR reporting the lowest agreement (Dice = 0.82), and TTA obtaining the highest score (Dice = 0.94).

CONCLUSIONS

Our outcomes highlight the importance of quantifying the segmentation uncertainty and that decision-makers can choose the approach most in line with the risk propensity degree required by the application and their policy.

Quantifying uncertainty in graph neural network explanations

In recent years, analyzing the explanation for the prediction of Graph Neural Networks (GNNs) has attracted increasing attention. Despite this progress, most existing methods do not adequately consider the inherent uncertainties stemming from the randomness of model parameters and graph data, which may lead to overconfidence and misguiding explanations. However, it is challenging for most of GNN explanation methods to quantify these uncertainties since they obtain the prediction explanation in a post-hoc and model-agnostic manner without considering the randomness of graph data and model parameters. To address the above problems, this paper proposes a novel uncertainty quantification framework for GNN explanations. For mitigating the randomness of graph data in the explanation, our framework accounts for two distinct data uncertainties, allowing for a direct assessment of the uncertainty in GNN explanations. For mitigating the randomness of learned model parameters, our method learns the parameter distribution directly from the data, obviating the need for assumptions about specific distributions. Moreover, the explanation uncertainty within model parameters is also quantified based on the learned parameter distributions. This holistic approach can integrate with any post-hoc GNN explanation methods. Empirical results from our study show that our proposed method sets a new standard for GNN explanation performance across diverse real-world graph benchmarks.

Trustworthy clinical AI solutions: A unified review of uncertainty quantification in Deep Learning models for medical image analysis

The full acceptance of Deep Learning (DL) models in the clinical field is rather low with respect to the quantity of high-performing solutions reported in the literature. End users are particularly reluctant to rely on the opaque predictions of DL models. Uncertainty quantification methods have been proposed in the literature as a potential solution, to reduce the black-box effect of DL models and increase the interpretability and the acceptability of the result by the final user. In this review, we propose an overview of the existing methods to quantify uncertainty associated with DL predictions. We focus on applications to medical image analysis, which present specific challenges due to the high dimensionality of images and their variable quality, as well as constraints associated with real-world clinical routine. Moreover, we discuss the concept of structural uncertainty, a corpus of methods to facilitate the alignment of segmentation uncertainty estimates with clinical attention. We then discuss the evaluation protocols to validate the relevance of uncertainty estimates. Finally, we highlight the open challenges for uncertainty quantification in the medical field.

Comparative evaluation of uncertainty estimation and decomposition methods on liver segmentation

PURPOSE

Deep neural networks need to be able to indicate error likelihood via reliable estimates of their predictive uncertainty when used in high-risk scenarios, such as medical decision support. This work contributes a systematic overview of state-of-the-art approaches for decomposing predictive uncertainty into aleatoric and epistemic components, and a comprehensive comparison for Bayesian neural networks (BNNs) between mutual information decomposition and the explicit modelling of both uncertainty types via an additional loss-attenuating neuron.

METHODS

Experiments are performed in the context of liver segmentation in CT scans. The quality of the uncertainty decomposition in the resulting uncertainty maps is qualitatively evaluated, and quantitative behaviour of decomposed uncertainties is systematically compared for different experiment settings with varying training set sizes, label noise, and distribution shifts.

RESULTS

Our results show the mutual information decomposition to robustly yield meaningful aleatoric and epistemic uncertainty estimates, while the activation of the loss-attenuating neuron appears noisier with non-trivial convergence properties. We found that the addition of a heteroscedastic neuron does not significantly improve segmentation performance or calibration, while slightly improving the quality of uncertainty estimates.

CONCLUSIONS

Mutual information decomposition is simple to implement, has mathematically pleasing properties, and yields meaningful uncertainty estimates that behave as expected under controlled changes to our data set. The additional extension of BNNs with loss-attenuating neurons provides no improvement in terms of segmentation performance or calibration in our setting, but marginal benefits regarding the quality of decomposed uncertainties.

A study of Bayesian deep network uncertainty and its application to synthetic CT generation for MR-only radiotherapy treatment planning

BACKGROUND

The use of synthetic computed tomography (CT) for radiotherapy treatment planning has received considerable attention because of the absence of ionizing radiation and close spatial correspondence to source magnetic resonance (MR) images, which have excellent tissue contrast. However, in an MR-only environment, little effort has been made to examine the quality of synthetic CT images without using the original CT images.

PURPOSE

To estimate synthetic CT quality without referring to original CT images, this study established the relationship between synthetic CT uncertainty and Bayesian uncertainty, and proposed a new Bayesian deep network for generating synthetic CT images and estimating synthetic CT uncertainty for MR-only radiotherapy treatment planning.

METHODS AND MATERIALS

A novel deep Bayesian network was formulated using probabilistic network weights. Two mathematical expressions were proposed to quantify the Bayesian uncertainty of the network and synthetic CT uncertainty, which was closely related to the mean absolute error (MAE) in Hounsfield Unit (HU) of synthetic CT. These uncertainties were examined to demonstrate the accuracy of representing the synthetic CT uncertainty using a Bayesian counterpart. We developed a hybrid Bayesian architecture and a new data normalization scheme, enabling the Bayesian network to generate both accurate synthetic CT and reliable uncertainty information when probabilistic weights were applied. The proposed method was evaluated in 59 patients (13/12/32/2 for training/validation/testing/uncertainty visualization) diagnosed with prostate cancer, who underwent same-day pelvic CT- and MR-acquisitions. To assess the relationship between Bayesian and synthetic CT uncertainties, linear and non-linear correlation coefficients were calculated on per-voxel, per-tissue, and per-patient bases. For accessing the accuracy of the CT number and dosimetric accuracy, the proposed method was compared with a commercially available atlas-based method (MRCAT) and a U-Net conditional-generative adversarial network (UcGAN).

RESULTS

The proposed model exhibited 44.33 MAE, outperforming UcGAN 52.51 and MRCAT 54.87. The gamma rate (2%/2 mm dose difference/distance to agreement) of the proposed model was 98.68%, comparable to that of UcGAN (98.60%) and MRCAT (98.56%). The per-patient and per-tissue linear correlation coefficients between the Bayesian and synthetic CT uncertainties ranged from 0.53 to 0.83, implying a moderate to strong linear correlation. Per-voxel correlation coefficients varied from -0.13 to 0.67 depending on the regions-of-interest evaluated, indicating tissue-dependent correlation. The R2 value for estimating MAE solely using Bayesian uncertainty was 0.98, suggesting that the uncertainty of the proposed model was an ideal candidate for predicting synthetic CT error, without referring to the original CT.

CONCLUSION

This study established a relationship between the Bayesian model uncertainty and synthetic CT uncertainty. A novel Bayesian deep network was proposed to generate a synthetic CT and estimate its uncertainty. Various metrics were used to thoroughly examine the relationship between the uncertainties of the proposed Bayesian model and the generated synthetic CT. Compared with existing approaches, the proposed model showed comparable CT number and dosimetric accuracies. The experiments showed that the proposed Bayesian model was capable of producing accurate synthetic CT, and was an effective indicator of the uncertainty and error associated with synthetic CT in MR-only workflows.

Nuclei instance segmentation from histopathology images using Bayesian dropout based deep learning

BACKGROUND

The deterministic deep learning models have achieved state-of-the-art performance in various medical image analysis tasks, including nuclei segmentation from histopathology images. The deterministic models focus on improving the model prediction accuracy without assessing the confidence in the predictions.

METHODS

We propose a semantic segmentation model using Bayesian representation to segment nuclei from the histopathology images and to further quantify the epistemic uncertainty. We employ Bayesian approximation with Monte-Carlo (MC) dropout during the inference time to estimate the model's prediction uncertainty.

RESULTS

We evaluate the performance of the proposed approach on the PanNuke dataset, which consists of 312 visual fields from 19 organ types. We compare the nuclei segmentation accuracy of our approach with that of a fully convolutional neural network, U-Net, SegNet, and the state-of-the-art Hover-net. We use F1-score and intersection over union (IoU) as the evaluation metrics. The proposed approach achieves a mean F1-score of 0.893 ± 0.008 and an IoU value of 0.868 ± 0.003 on the test set of the PanNuke dataset. These results outperform the Hover-net, which has a mean F1-score of 0.871 ± 0.010 and an IoU value of 0.840 ± 0.032.

CONCLUSIONS

The proposed approach, which incorporates Bayesian representation and Monte-Carlo dropout, demonstrates superior performance in segmenting nuclei from histopathology images compared to existing models such as U-Net, SegNet, and Hover-net. By considering the epistemic uncertainty, our model provides a more reliable estimation of the prediction confidence. These findings highlight the potential of Bayesian deep learning for improving medical image analysis tasks and can contribute to the development of more accurate and reliable computer-aided diagnostic systems.

CAM-VT: A Weakly supervised cervical cancer nest image identification approach using conjugated attention mechanism and visual transformer

Cervical cancer is the fourth most common cancer among women, and cytopathological images are often used to screen for this cancer. However, manual examination is very troublesome and the misdiagnosis rate is high. In addition, cervical cancer nest cells are denser and more complex, with high overlap and opacity, increasing the difficulty of identification. The appearance of the computer aided automatic diagnosis system solves this problem. In this paper, a weakly supervised cervical cancer nest image identification approach using Conjugated Attention Mechanism and Visual Transformer (CAM-VT), which can analyze pap slides quickly and accurately. CAM-VT proposes conjugated attention mechanism and visual transformer modules for local and global feature extraction respectively, and then designs an ensemble learning module to further improve the identification capability. In order to determine a reasonable interpretation, comparative experiments are conducted on our datasets. The average accuracy of the validation set of three repeated experiments using CAM-VT framework is 88.92%, which is higher than the optimal result of 22 well-known deep learning models. Moreover, we conduct ablation experiments and extended experiments on Hematoxylin and Eosin stained gastric histopathological image datasets to verify the ability and generalization ability of the framework. Finally, the top 5 and top 10 positive probability values of cervical nests are 97.36% and 96.84%, which have important clinical and practical significance. The experimental results show that the proposed CAM-VT framework has excellent performance in potential cervical cancer nest image identification tasks for practical clinical work.

Robust Motor Imagery Tasks Classification Approach Using Bayesian Neural Network

The development of Brain-Computer Interfaces based on Motor Imagery (MI) tasks is a relevant research topic worldwide. The design of accurate and reliable BCI systems remains a challenge, mainly in terms of increasing performance and usability. Classifiers based on Bayesian Neural Networks are proposed in this work by using the variational inference, aiming to analyze the uncertainty during the MI prediction. An adaptive threshold scheme is proposed here for MI classification with a reject option, and its performance on both datasets 2a and 2b from BCI Competition IV is compared with other approaches based on thresholds. The results using subject-specific and non-subject-specific training strategies are encouraging. From the uncertainty analysis, considerations for reducing computational cost are proposed for future work.

SkiNet: A deep learning framework for skin lesion diagnosis with uncertainty estimation and explainability

Skin cancer is considered to be the most common human malignancy. Around 5 million new cases of skin cancer are recorded in the United States annually. Early identification and evaluation of skin lesions are of great clinical significance, but the disproportionate dermatologist-patient ratio poses a significant problem in most developing nations. Therefore a novel deep architecture, named as SkiNet, is proposed to provide faster screening solution and assistance to newly trained physicians in the process of clinical diagnosis of skin cancer. The main motive behind SkiNet's design and development is to provide a white box solution, addressing a critical problem of trust and interpretability which is crucial for the wider adoption of Computer-aided diagnosis systems by medical practitioners. The proposed SkiNet is a two-stage pipeline wherein the lesion segmentation is followed by the lesion classification. Monte Carlo dropout and test time augmentation techniques have been employed in the proposed method to estimate epistemic and aleatoric uncertainty. A novel segmentation model named Bayesian MultiResUNet is used to estimate the uncertainty on the predicted segmentation map. Saliency-based methods like XRAI, Grad-CAM and Guided Backprop are explored to provide post-hoc explanations of the deep learning models. The ISIC-2018 dataset is used to perform the experimentation and ablation studies. The results establish the robustness of the proposed model on the traditional benchmarks while addressing the black-box nature of such models to alleviate the skepticism of medical practitioners by incorporating transparency and confidence to the model's prediction.

An Efficient and Uncertainty-Aware Decision Support System for Disaster Response Using Aerial Imagery

Efficient and robust search and rescue actions are always required when natural or technical disasters occur. Empowered by remote sensing techniques, building damage assessment can be achieved by fusing aerial images of pre- and post-disaster environments through computational models. Existing methods pay over-attention to assessment accuracy without considering model efficiency and uncertainty quantification in such a life-critical application. Thus, this article proposes an efficient and uncertain-aware decision support system (EUDSS) that evolves the recent computational models into an efficient decision support system, realizing the uncertainty during building damage assessment (BDA). Specifically, a new efficient and uncertain-aware BDA integrates the recent advances in computational models such as Fourier attention and Monte Carlo Dropout for uncertainty quantification efficiently. Meanwhile, a robust operation (RO) procedure is designed to invite experts for manual reviews if the uncertainty is high due to external factors such as cloud clutter and poor illumination. This procedure can prevent rescue teams from missing damaged houses during operations. The effectiveness of the proposed system is demonstrated on a public dataset from both quantitative and qualitative perspectives. The solution won the first place award in International Overhead Imagery Hackathon.

Towards a safe and efficient clinical implementation of machine learning in radiation oncology by exploring model interpretability, explainability and data-model dependency

The interest in machine learning (ML) has grown tremendously in recent years, partly due to the performance leap that occurred with new techniques of deep learning, convolutional neural networks for images, increased computational power, and wider availability of large datasets. Most fields of medicine follow that popular trend and, notably, radiation oncology is one of those that are at the forefront, with already a long tradition in using digital images and fully computerized workflows. ML models are driven by data, and in contrast with many statistical or physical models, they can be very large and complex, with countless generic parameters. This inevitably raises two questions, namely, the tight dependence between the models and the datasets that feed them, and the interpretability of the models, which scales with its complexity. Any problems in the data used to train the model will be later reflected in their performance. This, together with the low interpretability of ML models, makes their implementation into the clinical workflow particularly difficult. Building tools for risk assessment and quality assurance of ML models must involve then two main points: interpretability and data-model dependency. After a joint introduction of both radiation oncology and ML, this paper reviews the main risks and current solutions when applying the latter to workflows in the former. Risks associated with data and models, as well as their interaction, are detailed. Next, the core concepts of interpretability, explainability, and data-model dependency are formally defined and illustrated with examples. Afterwards, a broad discussion goes through key applications of ML in workflows of radiation oncology as well as vendors' perspectives for the clinical implementation of ML.

Automatic brain segmentation in preterm infants with post-hemorrhagic hydrocephalus using 3D Bayesian U-Net

Post‐hemorrhagic hydrocephalus (PHH) is a severe complication of intraventricular hemorrhage (IVH) in very preterm infants. PHH monitoring and treatment decisions rely heavily on manual and subjective two‐dimensional measurements of the ventricles. Automatic and reliable three‐dimensional (3D) measurements of the ventricles may provide a more accurate assessment of PHH, and lead to improved monitoring and treatment decisions. To accurately and efficiently obtain these 3D measurements, automatic segmentation of the ventricles can be explored. However, this segmentation is challenging due to the large ventricular anatomical shape variability in preterm infants diagnosed with PHH. This study aims to (a) propose a Bayesian U‐Net method using 3D spatial concrete dropout for automatic brain segmentation (with uncertainty assessment) of preterm infants with PHH; and (b) compare the Bayesian method to three reference methods: DenseNet, U‐Net, and ensemble learning using DenseNets and U‐Nets. A total of 41 T₂‐weighted MRIs from 27 preterm infants were manually segmented into lateral ventricles, external CSF, white and cortical gray matter, brainstem, and cerebellum. These segmentations were used as ground truth for model evaluation. All methods were trained and evaluated using 4‐fold cross‐validation and segmentation endpoints, with additional uncertainty endpoints for the Bayesian method. In the lateral ventricles, segmentation endpoint values for the DenseNet, U‐Net, ensemble learning, and Bayesian U‐Net methods were mean Dice score = 0.814 ± 0.213, 0.944 ± 0.041, 0.942 ± 0.042, and 0.948 ± 0.034 respectively. Uncertainty endpoint values for the Bayesian U‐Net were mean recall = 0.953 ± 0.037, mean  negative predictive value = 0.998 ± 0.005, mean accuracy = 0.906 ± 0.032, and mean AUC = 0.949 ± 0.031. To conclude, the Bayesian U‐Net showed the best segmentation results across all methods and provided accurate uncertainty maps. This method may be used in clinical practice for automatic brain segmentation of preterm infants with PHH, and lead to better PHH monitoring and more informed treatment decisions.

Bacterial species identification using MALDI-TOF mass spectrometry and machine learning techniques: A large-scale benchmarking study

Today machine learning methods are commonly deployed for bacterial species identification using MALDI-TOF mass spectrometry data. However, most of the studies reported in literature only consider very traditional machine learning methods on small datasets that contain a limited number of species. In this paper we present benchmarking results on an unprecedented scale for a wide range of machine learning methods, using datasets that contain almost 100,000 spectra and more than 1000 different species. The size and the diversity of the data allow to compare three important identification scenarios that are often not distinguished in literature, i.e., identification for novel biological replicates, novel strains and novel species that are not present in the training data. The results demonstrate that in all three scenarios acceptable identification rates are obtained, but the numbers are typically lower than those reported in studies with a more limited analysis. Using hierarchical classification methods, we also demonstrate that taxonomic information is in general not well preserved in MALDI-TOF mass spectrometry data. For the novel species scenario, we apply for the first time neural networks with Monte Carlo dropout, which have shown to be successful in other domains, such as computer vision, for the detection of novel species.

Monte Carlo Dropout for Uncertainty Estimation and Motor Imagery Classification

Motor Imagery (MI)-based Brain-Computer Interfaces (BCIs) have been widely used as an alternative communication channel to patients with severe motor disabilities, achieving high classification accuracy through machine learning techniques. Recently, deep learning techniques have spotlighted the state-of-the-art of MI-based BCIs. These techniques still lack strategies to quantify predictive uncertainty and may produce overconfident predictions. In this work, methods to enhance the performance of existing MI-based BCIs are proposed in order to obtain a more reliable system for real application scenarios. First, the Monte Carlo dropout (MCD) method is proposed on MI deep neural models to improve classification and provide uncertainty estimation. This approach was implemented using Shallow Convolutional Neural Network (SCNN-MCD) and with an ensemble model (E-SCNN-MCD). As another contribution, to discriminate MI task predictions of high uncertainty, a threshold approach is introduced and tested for both SCNN-MCD and E-SCNN-MCD approaches. The BCI Competition IV Databases 2a and 2b were used to evaluate the proposed methods for both subject-specific and non-subject-specific strategies, obtaining encouraging results for MI recognition.

Bayesian Fully Convolutional Networks for Brain Image Registration

The purpose of medical image registration is to find geometric transformations that align two medical images so that the corresponding voxels on two images are spatially consistent. Nonrigid medical image registration is a key step in medical image processing, such as image comparison, data fusion, target recognition, and pathological change analysis. Existing registration methods only consider registration accuracy but largely neglect the uncertainty of registration results. In this work, a method based on the Bayesian fully convolutional neural network is proposed for nonrigid medical image registration. The proposed method can generate a geometric uncertainty map to calculate the uncertainty of registration results. This uncertainty can be interpreted as a confidence interval, which is essential for judging whether the source data are abnormal. Moreover, the proposed method introduces group normalization, which is conducive to the network convergence of the Bayesian neural network. Some representative learning-based image registration methods are compared with the proposed method on different image datasets. Experimental results show that the registration accuracy of the proposed method is better than that of the methods, and its antifolding performance is comparable to that of fast image registration and VoxelMorph. Furthermore, the proposed method can evaluate the uncertainty of registration results.

Uncertainty quantification in skin cancer classification using three-way decision-based Bayesian deep learning

Accurate automated medical image recognition, including classification and segmentation, is one of the most challenging tasks in medical image analysis. Recently, deep learning methods have achieved remarkable success in medical image classification and segmentation, clearly becoming the state-of-the-art methods. However, most of these methods are unable to provide uncertainty quantification (UQ) for their output, often being overconfident, which can lead to disastrous consequences. Bayesian Deep Learning (BDL) methods can be used to quantify uncertainty of traditional deep learning methods, and thus address this issue. We apply three uncertainty quantification methods to deal with uncertainty during skin cancer image classification. They are as follows: Monte Carlo (MC) dropout, Ensemble MC (EMC) dropout and Deep Ensemble (DE). To further resolve the remaining uncertainty after applying the MC, EMC and DE methods, we describe a novel hybrid dynamic BDL model, taking into account uncertainty, based on the Three-Way Decision (TWD) theory. The proposed dynamic model enables us to use different UQ methods and different deep neural networks in distinct classification phases. So, the elements of each phase can be adjusted according to the dataset under consideration. In this study, two best UQ methods (i.e., DE and EMC) are applied in two classification phases (the first and second phases) to analyze two well-known skin cancer datasets, preventing one from making overconfident decisions when it comes to diagnosing the disease. The accuracy and the F1-score of our final solution are, respectively, 88.95% and 89.00% for the first dataset, and 90.96% and 91.00% for the second dataset. Our results suggest that the proposed TWDBDL model can be used effectively at different stages of medical image analysis.