CytoDNNEnsemble: combining cell-level and distribution-level representations for flow-cytometry classification

Flow cytometry is a cornerstone diagnostic technique, but manual analysis is laborious, subjective, and prone to inter-operator variability. We propose CytoDNNEnsemble , a compact deep ensemble model for automated analysis of high-dimensional flow cytometry data. The architecture consists of two smaller, independent subnetworks that process blood samples from complementary perspectives. This design may reduce reliance on expert gating and improve diagnostic consistency. Across four public benchmarks, CDE achieves higher mean accuracy and AUC than the published comparison values on AML and HEUvsUE, while performance on HVTN and NK-Cell is comparable to or below the strongest reported baseline. These findings indicate that the benefit of feature fusion is dataset dependent and warrants validation under fully controlled benchmark implementations and external cohorts.