A Dirichlet Distribution-Based Trust-Adaptive Ensemble Approach for Pneumonia Classification from Chest X-Ray Images

Pneumonia diagnosis from chest radiographs is hindered by AI model variability and limited decision transparency. We present a deployment-ready, abstaining Dirichlet-evidence ensemble for paediatric CXR triage that elevates “Indeterminate” to a first-class outcome and implements dynamic, trust-adaptive weighting. Eight diverse pre-trained models each provide binary predictions, which are converted to three-way CXR supports, scaled by ongoing trust scores, and fused using a Dirichlet-evidence mechanism. Selective gates on class separation and pooled evidence allow the ensemble to defer, with auditable reasoning, when the decision is uncertain. On a 200-image paediatric hold-out, simple majority voting achieved 94.5% accuracy but dropped to 83.0% when two training models were intentionally inverted. At a fixed operating point, the proposed ensemble maintained zero errors (100% accuracy) on decided cases in both settings, abstaining on only 1-2% of studies instead of issuing incorrect labels. Unlike prior work, our system tightly integrates uncertainty, selective prediction, and per-case logging for clinical governance. This framework advances safe, interpretable automation for paediatric pneumonia assessment.