Making hematology AI reliable: adversarial training for photometric and 4D variation

Medical-imaging AI often reports high accuracy under nominal conditions, while most “adversarial” studies target imperceptible Lp-bounded noise (FGSM/PGD). In digital microscopy, failures instead arise from acquisition shifts (brightness, contrast, channel/intensity). We close this gap by evaluating and hardening blood-smear classifiers against photometric and structured 4D perturbations to achieve clinically realistic robustness beyond pixel-bounded attacks.We constructed a shift-robustness evaluation encompassing Original, Brightness, Contrast, and 4D-change (additional channel/intensity) conditions. A standard blood-cell classifier was trained under two regimes: (I) baseline (no adversarial training) and (II) modality-aware adversarial training that injects photometric and 4D perturbations while preserving clean-image performance via balanced sampling and routine augmentations. Performance was assessed by overall and per-class accuracy, confusion matrices, and prediction flip counts across eight cell types; we additionally recommend calibration analyses (ECE, Brier score). This work motivates modality-aware adversarial training as a practical path to safer hematology AI and offers a reusable evaluation template (conditions, metrics, plots) for microscopy systems beyond blood-cell classification.