Automated classification of white blood cell (WBC) images may support cell-morphology workflows, but high accuracy alone does not establish clinical readiness. We developed an Uncertainty-Guided Deep Ensemble (UGDE) combining MobileNetV3-Large, EfficientNet-B0, and ShuffleNetV2 with Monte Carlo dropout, temperature scaling, split conformal prediction, and Grad-CAM. The primary dataset contained 16,027 single-cell images from 78 patients (33 with leukemia and 45 without leukemia). All stated development partitions were grouped by patient, and augmentation was restricted to the proper training subset. On the held-out internal test set (3,205 images from 16 patients), UGDE achieved approximately 99% image-level accuracy and macro-F1 of 0.99. Expected calibration error was 2.3% before and 0.8% after temperature scaling. Internal empirical conformal coverage was 94.2%, 89.7%, and 79.7% at nominal coverage levels of 95%, 90%, and 80%, respectively; these slight shortfalls are reported without claiming that the targets were met. External evaluation used 2000 Raabin-WBC images after mapping the nine model outputs to five shared leukocyte categories and yielded 94% descriptive accuracy. Grad-CAM examples were used as post-hoc visualizations and were not interpreted as causal explanations. These retrospective results support further evaluation of uncertainty-aware WBC image classification, but they do not demonstrate performance on whole slides, automated differential counts, prospective patients, or clinical workflows.
Federated learning enables cross-institutional clinical-genetics analytics without sharing raw patient data; yet, no existing framework integrates the full cryptographic stack-differential privacy, homomorphic encryption, and blockchain auditing-required for regulatory deployment. Here, we present FFT-GA-BC, a secure-by-construction pipeline combining a bit-flip genetic algorithm for binary feature selection across 136-dimensional genomic representations, a multi-head self-attention encoder, population-stratified federated averaging, (ε,δ)-differential privacy via the analytic Gaussian mechanism, Cheon-Kim-Kim-Song homomorphic encryption for cipher-domain weight aggregation, and a Practical Byzantine Fault Tolerance (PBFT)-consensus blockchain with SHA-256 Merkle-root commitments per federation round. We evaluate the framework under a simulated multi-institutional deployment scenario using a 950-record synthetic benchmark whose schema mirrors the Punjab Thalassemia and Genetic Disorders Programme Lahore registry, a South Asian clinical-genetics registry operating across six ancestry strata federated across a simulated 12-hospital, five-continent topology, and internally externally validated via a leave-one-population-out proxy across all six ancestry cohorts. Under these conditions, differential privacy imposes less than a 2.7 percentage-point utility loss across the clinically relevant privacy-budget range, with ε = 1 recommended as the clinical operating point (Section 5.8), and fully encrypted federated training matches centralised pooled training within ±1.6 percentage points across five independent splits. The blockchain audit layer scales to 192 nodes at 1.3 s of per-round latency with on-chain audit costs below USD 350 per block, maintaining an O(log K) Merkle-inclusion proof complexity per query. Information-theoretic profiling of the benchmark confirms that labels carry at most 0.036 bits of mutual information against a target entropy of 2.0 bits, demonstrating that the pipeline's diagnostic output faithfully reflects data structure rather than architectural artefacts, a critical validation property for clinical deployment. To the best of our knowledge, based on a systematic literature search (PubMed, IEEE Xplore, arXiv; terms: federated learning + genomics + privacy + blockchain; June 2026), FFT-GA-BC provides the first fully specified, reproducible, and externally validated end-to-end secure federated genomics pipeline architecturally and infrastructurally ready for multi-hospital deployment, pending real-cohort clinical validation.