Enhancing Diagnostic in 3D COVID-19 Pneumonia CT-scans through Explainable Uncertainty Bayesian Quantification Journal: Artificial Intelligence in Medicine

Juan Manuel Liscano Fierro, Héctor J. Hortúa · 2025

Read Enhancing Diagnostic in 3D COVID-19 Pneumonia CT-scans through Explainable Uncertainty Bayesian Quantification Journal: Artificial Intelligence in Medicine by Juan Manuel Liscano Fierro, Héctor J. Hortúa (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2501-10770.txt.

Introduction

Juan Manuel Liscano Fierro and Héctor Hortúa apply Bayesian neural networks to reading 3D CT scans for COVID-19 pneumonia. Standard networks give a single answer with no sense of their own doubt, which weakens clinical use. The team compared plain networks with Bayesian versions that report uncertainty, tuning both heavily.

Lightweight designs reached the best accuracy, at ninety-six percent. Their Bayesian forms, built with a multiplied normalizing flow method, kept that accuracy while giving well-calibrated doubt estimates. A 3D display based on SHAP values shows which scan regions drove each call. The authors argue that doubt plus explanation makes for safer medical image decisions.

Catalog record

CREATORJuan Manuel Liscano Fierro, Héctor J. Hortúa
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE69 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2501-10770

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Enhancing Diagnostic in 3D COVID-19 Pneumonia CT-scans through Explainable Uncertainty Bayesian Quantification Journal: Artificial Intelligence in Medicine Juan Manuel Liscano Fierro, Héctor J. Hortúa Abstract Accurately classifying COVID-19 pneumonia in 3D CT scans remains a significant challenge in the field of medical image analysis. Although deterministic neural networks have shown promising results in this area, they provide only point estimates outputs yielding poor diagnostic in clinical decision-making. In this paper, we explore the use of Bayesian neural networks for classifying COVID-19 pneumonia in 3D CT scans providing uncertainties in their predictions. We compare deterministic networks and their Bayesian counterpart, enhancing the decision-making accuracy under uncertainty information. Remarkably, our findings reveal that lightweight architectures achieve the highest accuracy of 96% after developing extensive hyperparameter tuning. Furthermore, the Bayesian counterpart of these architectures via Multiplied Normalizing Flow technique kept a similar performance along with calibrated uncertainty estimates. Finally, we have developed a 3D-visualization approach to explain the neural network outcomes based on SHAP values. We conclude that explainability along with uncertainty quantification will offer better clinical decisions in medical image analysis, contributing to ongoing efforts for improving the diagnosis and treatment of COVID-19 pneumonia. 1 Introduction The rapid integration of artificial intelligence (AI) into the medical field, particularly in the domain of clinical imaging, has led to significant advancements in diagnostic capabilities Shenavarmasouleh et al. 2023 . However, despite these technological strides, one of the major challenges that persists is ensuring the reliability and transparency of AI-driven diagnostic tools, especially in high-stakes environments like healthcare Arsenos et al. 2022 ; Arsenos et al. 2023 . In the context of COVID-19 pneumonia, accurate classification of 3D CT scans is crucial for effective patient management, but existing methods often lack the ability to adequately quantify the uncertainty in their predictions—a critical factor in clinical decision-making Li et al. 2024 . Traditional neural networks, widely employed for medical image analysis, have achieved impressive accuracy in various diagnostic tasks Kollias et al. 2020 ; Kollias et al. 2021 ; Kollias et al. 2023 ; Hou et al. 2022 . Nonetheless, these models typically operate in a deterministic manner, providing single-point estimates without indicating the level of confidence in their predictions. This absence of uncertainty quantification can lead to overconfident and potentially misleading decisions, which is particularly concerning in medical settings where the cost of error is high Zahari et al. 2024 . To address this limitation, recent research has explored the use of Bayesian neural networks, which offer a probabilistic framework that not only produces predictions but also quantifies the associated uncertainty. Several studies have proposed methods such as Monte Carlo Dropout Zhao et al. 2024 Bargagna et al. 2023a Zou et al. 2023 and Variational Inference Hu et al. 2023 to approximate Bayesian inference in neural networks, demonstrating their potential to improve the interpretability and reliability of AI models in medical imaging. However, these approaches often involve trade-offs, such as increased computational complexity and potential overestimation of uncertainty, which require further investigation Guo et al. 2017 ; Kristiadi et al. 2020 . In this research, we aim to compare the effectiveness of deterministic neural networks and Bayesian neural networks in the classification of COVID-19 pneumonia using 3D CT scans. Our approach focuses on leveraging the strengths of Bayesian methods to enhance uncertainty quantification, thereby refining the clinical decision-making process. The main objective of this study is to determine whether Bayesian neural networks can provide more reliable and interpretable predictions compared to traditional neural networks, particularly in the context of medical imaging. The authors in Garcia-Farieta et al. 2024 have claimed that Bayesian neural networks with Multiplicative Normalising Flows(MNF) 3D-layers demonstrate superior performance in terms of uncertainty quantification, providing more actionable insights. These findings highlight the potential of Bayesian approaches to not only improve diagnostic accuracy but also to offer a more nuanced understanding of the confidence in AI-driven diagnoses, ultimately contributing to safer and more effective clinical practices. Finally, we have developed a 3D- explainable routine based on Shapley values Lundberg and Lee 2017b to understand deep learning models decision by highlighting the features deemed relevant by the model. It allowed us to observe regions on the scans relevant for classifying the degree of COVID-19 pneumonia. The paper is developed as follows, in section 3 we explain the concepts behind of Bayesian Neural Networks and describe briefly some of the main techniques used in this paper for adding uncertainty in the models, in section 4 we have a glimpse about calibration methods and metrics for evaluating the accuracy in the uncertainty estimates and section 5 we provide information about the metrics used for evaluating the performance of the model predictions. Section 6 describes the dataset used and the methodology is commented in section 7 . Results are shown in section 8 and finally the discussion along with conclusions are mentioned in section 9 . 2 Related work The need to enhance the reliability and trustworthiness of AI models in healthcare, especially in medical image analysis, is a critical challenge An et al. 2021 ; Zhao et al. 2024 ; Bargagna et al. 2023b . While AI-driven diagnostic tools have made significant progress, the persistent issue of uncertainty and potential unreliability in model predictions remains a concern for clinical decision-making Zou et al. 2023 ; Hu et al. 2023 ; Abboud et al. 2024 . Bayesian neural networks (BNNs) offer a promising solution by providing probabilistic interpretations of predictions, allowing clinicians to better assess prediction confidence Zhang et al. 2024 . The COVID-19 pandemic has brought renewed attention to medical image analysis, particularly in the classification of COVID-19 pneumonia using chest CT scans. Deep learning techniques, such as convolutional neural networks (CNNs), have shown great promise in this area, with successful applications in detecting intracranial hemorrhages Sharrock et al. 2021 ; Chang et al. 2018 and diagnosing COVID-19-related lung conditions Chen et al. 2021 . However, deterministic neural networks, while effective, do not account for the uncertainty inherent in medical imaging, a limitation that BNNs can overcome. Recent studies underscore the potential of deep learning in improving diagnostic accuracy and streamlining clinical workflows Lundervold and Lundervold 2019 ; Shen et al. 2017 . For instance, McKinney et al. McKinney et al. 2020 demonstrated that deep learning algorithms could detect breast cancer with accuracy comparable to experienced radiologists. Similarly, Ardila et al. Ardila et al. 2019 highlighted the effectiveness of AI models in detecting lung cancer from CT scans, showing potential for enhancing global cancer screening. Among the deep learning techniques, BNNs are particularly valuable in medical imaging due to their ability to quantify uncertainty, a crucial factor in clinical decision-making. This uncertainty quantification enables more informed judgments about the reliability of diagnoses and supports more nuanced treatment decisions Papamarkou et al. 2024 . 3 Bayesian Neural Networks Bayesian neural networks (BNNs) are a type of neural network that incorporates Bayesian probability theory into the training and inference process. Unlike deterministic neural networks, which generate a fixed prediction given a set of inputs, BNNs generate a probability distribution over possible predictions ( Chai 2018 ) . In Chai 2018 discusses the importance of decomposing predictive uncertainty by arguing that random and epistemic uncertainties tell us about different

Read the complete text