Bayesian Deep Learning for Uncertainty-Aware Cancer Diagnosis in Medical AI Systems

Despite efforts to prevent its development and
eradicate it, cancer remains among the top ten leading causes of deaths around the world and therefore the need for intelligent and clinically configurable diagnostic systems that can support the early detection and clinical decision making. Deep learning methods have achieved an outstanding performance in analyzing medical images, but current models only make a deterministic prediction, and they do not quantify the uncertainty in their prediction, which does not allow them to be trusted in realclinical contexts. The biggest challenge is conventional AI systems
struggle to handle the ambiguity, noise, and variation in medical data and deliver confidence scores. This study is motivated by the need for a novel Bayesian Multi-Modal Evidential Deep Learning Network (BMEDL-Net) to tackle this challenge in the field of uncertainty-aware cancer diagnosis. The framework contains histopathology images, radiological scans, and clinical information that are integrated together employing Bayesian feature extraction, variational representation learning, evidential
multi-modal fusion, two uncertainty estimation, and explainable decision support mechanisms. A multi-modal cancer dataset was used for experimental evaluation, which was carried out in a high-performance computing environment. The results indicate that the model successfully met the requirements with an accuracy of 98.76%, precision of 98.42%, recall of 98.15%, F1- score of 98.28%, and AUC of 99.12%, and an Expected Calibration Error of 2.14% and uncertainty estimation quality of
97.46%. Results are far better than current State of the Art (SoA) CNN, Transformer and Bayesian based methods. The results
show that the proposed framework significantly improves the
diagnostic reliability, interpretability and clinical trust,
highlighting its potential use in developing the next generation of
intelligent healthcare systems and precision oncology.