Enhancing Diagnostic in 3D COVID-19 Pneumonia CT-scans through Explainable Uncertainty Bayesian Quantification

TL;DR

This study tackles the reliability gap in AI-based 3D CT COVID-19 pneumonia classification by comparing deterministic CNNs with Bayesian neural networks that quantify predictive uncertainty. It combines variational inference and multiplicative normalising flows (MNF), including a voxel-grid extension for 3D convolutions, and evaluates calibration via reliability diagrams and ECE, alongside SHAP-based 3D explainability. Key findings show that lightweight deterministic models with careful hyperparameter tuning can reach up to $\approx96\%$ accuracy, while MNF-based BNNs achieve similar performance with calibrated uncertainty and informative prediction intervals. SHAP visualisations further provide voxel-level explanations, enhancing interpretability and potential clinical trust, suggesting uncertainty-aware, explainable AI can support safer decision-making in COVID-19 pneumonia diagnosis from 3D CT images.

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.

Figures (13)

• Figure 1: Metrics evaluated in the first group of models. We observe that DNN-W4-V1 outperforms in most of the metrics.
• Figure 2: Contrast of metrics evaluated in deterministic models before and after optimizing hyper-parameters.
• Figure 3: Contrast of metrics evaluated in all deterministic models.
• Figure 4: Performance of different BNNs methods based on metrics score. We report that MNF technique outperforms the rest of the models.
• Figure 8: 95% confidence intervals for estimated model probabilities for a deterministic and a Bayesian model, respectively. These plots show the uncertainty in predictions for a CT scan image, where the deterministic model provides fixed probability estimates, while the Bayesian model incorporates uncertainty, offering a range of possible outcomes through the confidence intervals.