Deep Learning-Based Smart AI System for Blood Cell Classification and Anemia Detection

This paper presented HemaVision, an open-access multimodal machine learning framework for preliminary anemia screening integrating peripheral blood smear photomicrographs and complete blood count parameters. By combining an anchor-free YOLOv8n cell enumerator, a MobileNetV3-Small cytomorphology classifier, a tabular Random Forest panel, and two multimodal fusion mechanisms (a 12-dimensional Stacking Ensemble and an experimental Gated Multimodal Unit), the framework demonstrates the feasibility of unified hematological screening. Through rigorous empirical audits, we uncovered and remediated critical methodological pitfalls common in medical AI: •  An exhaustive difference-hash and cosine similarity audit of 837 microscopy images identified 27 candidate duplicate pairs and 14 cross-split contaminated pairs in standard baselines. Reorganizing the dataset into 811 atomic clusters resolved data leakage, establishing a validated smear classification performance of 75.19% accuracy and 0.8167 ROC-AUC. •  An empirical target leakage audit revealed that standard CBC models achieve 99.51% to 100.0% accuracy through trivial memorization of the WHO hemoglobin threshold. A genuine screening model strictly omitting hemoglobin achieved 83.56% accuracy and 0.8974 ROC-AUC by learning biological surrogate markers in erythrocyte cytometric parameters (MCV, MCH, MCHC). •  Multi-paradigm multimodal integration demonstrated high synthetic test performance (99.44% Stacking Ensemble, 97.50% GMU), while mask-aware gating provided crash-safe operation across missing modalities. Crucially, we emphasize that high multimodal performance on synthetically paired datasets does not establish clinical diagnostic efficacy on human patients. HemaVision is an investigational research prototype, providing a transparent, scientifically validated baseline for future prospective clinical research in AI-assisted hematology.

1. World Health Organization, Haemoglobin concentrations for the diagnosis of anaemia and assessment of severity, Vitamin and Mineral Nutrition Information System, Geneva: World Health Organization, 2011 (WHO/NMH/NHD/MNM/11.1). 2. N. J. Kassebaum, R. Jasrasaria, M. Naghavi, et al., "A systematic analysis of global anemia burden from 1990 to 2010," Blood, vol. 123, no. 5, pp. 615–624, 2014, doi: 10.1182/blood-2013-06-508325. 3. J. A. Brittin and G. Brecher, "Instrumentation and automation in hematology," Prog. Hematol., vol. 7, pp. 299–340, 1971. 4. B. J. Bain, Blood Cells: A Practical Guide, 5th ed. Chichester, UK: Wiley-Blackwell, 2015. 5. G. L. Kumar and D. L. Bailey, Education Guide: Immunohistochemical Staining Methods, 6th ed. Carpinteria, CA: Dako North America, 2013. 6. A. Howard, M. Sandler, G. Chu, et al., "Searching for MobileNetV3," in Proc. IEEE/CVF Int. Conf. Comput. Vis. (ICCV), 2019, pp. 1314–1324, doi: 10.1109/ICCV.2019.00140. 7. L. Breiman, "Random Forests," Mach. Learn., vol. 45, no. 1, pp. 5–32, 2001, doi: 10.1023/A:1010933404324. 8. Shenggan, "BCCD (Blood Cell Count and Detection) Dataset," MIT License, 2019. [Online]. Available: https://github.com/Shenggan/BCCD_Dataset 9. A. Acevedo, A. Merino, L. Alférez, et al., "A dataset of microscopic peripheral blood cell images for development of automatic recognition systems," Data in Brief, vol. 30, Art. no. 105474, 2020, doi: 10.1016/j.dib.2020.105474. 10. G. Jocher, A. Chaurasia, and J. Qiu, "Ultralytics YOLOv8," Version 8.1.0, 2023. [Online]. Available: https://github.com/ultralytics/ultralytics