C-UQ: Conflict-based uncertainty quantification-A case study in lung cancer classification

Uncertainty quantification is crucial in deep learning, especially in medical diagnostics, to measure model prediction confidence and ensure reliable clinical decisions. This study introduces a novel conflict-based uncertainty quantification approach, applied as a case study in lung cancer classification, leveraging Dempster-Shafer Theory in conjunction with Deep Ensemble methods. The proposed method aggregates predictions from multiple neural network models using conflict as an uncertainty measure. By converting softmax outputs into Basic Belief Assignments and applying the rule of combination, this conflict-based method effectively quantifies uncertainty: high conflict values indicate predictions requiring expert review, and low values are considered reliable. Evaluations on the LIDC-IDRI dataset and additional 3D biomedical datasets show that the proposed method achieved high accuracy (0.957) and U_Recall (0.819) for lung classification. The sensitivity analysis further revealed that increasing the ensemble size enhanced performance even though the computational demands may challenge real-time applications. In contrast, the entropy-based smoothing effect limits the accuracy improvement of traditional Deep Ensemble methods. In addition, Out-of-Distribution detection with the proposed method achieved AUC scores up to 0.864 across various datasets. Future work will focus on optimising efficiency and exploring alternative Dempster-Shafer Theory combination rules and hybrid models.