A lens-free wide-field microfluidic platform for blood cell counting: comparison of deterministic and deep learning methods
Microfluidics and Nanofluidics, cilt.30, sa.11, 2026 (SCI-Expanded, Scopus)
- Yayın Türü: Makale / Tam Makale
- Cilt numarası: 30 Sayı: 11
- Basım Tarihi: 2026
- Doi Numarası: 10.1007/s10404-026-02937-5
- Dergi Adı: Microfluidics and Nanofluidics
- Derginin Tarandığı İndeksler: Science Citation Index Expanded (SCI-EXPANDED), Scopus, Compendex, INSPEC, Natural Science Collection (ProQuest), Health Research Premium Collection (ProQuest), Materials Science & Engineering Collection (ProQuest), Pharma Collection (ProQuest), Technology Collection (ProQuest)
- Anahtar Kelimeler: Deep learning detection, Deterministic image processing, Lens-free holographic imaging, Microfluidics, Point-of-care diagnostics, Red blood cell counting
- Ankara Üniversitesi Adresli: Evet
Özet
Complete blood count (CBC) testing is essential for clinical decision-making, but conventional analyzers are difficult to deploy in resource-limited and point-of-care settings. We present an in-house-developed, end-to-end lens-free wide-field microfluidic platform comprising an imaging module, a disposable laminated cartridge, and analysis software for red blood cell (RBC) enumeration from raw sensor-plane holograms. Cartridge characterization showed an 87.5% fabrication yield, 80.0% complete filling among successfully fabricated cartridges, and a mean filling time of 3.6 s. A deterministic image-processing pipeline and a YOLOv12 deep-learning detector were compared on the same holograms. Performance was evaluated using microscope-guided manual annotations of RBCs and visually confounding non-cell patterns directly on the holograms. Because true negatives are not naturally defined in instance-level detection, specificity is often omitted. Treating these patterns as hard negatives enabled specificity and accuracy to be reported alongside sensitivity, precision, and F1-score for both methods. Inter-observer agreement was high, with intraclass correlation coefficients of 0.962 for RBCs and 0.860 for hard negatives. YOLOv12 outperformed the deterministic pipeline (F1-score 0.882 vs. 0.834) and showed closer agreement with the clinical CBC reference (mean absolute percentage error [MAPE], 4.7% vs. 7.7%); manual counting yielded a MAPE of 1.3%. Without retraining, YOLOv12 retained lower MAPE on two additional independent samples (4.9% vs. 8.2% and 5.7% vs. 7.1%). These proof-of-concept findings support the technical feasibility of the lens-free microfluidic workflow and demonstrate the value of specificity-aware evaluation for distinguishing missed cells from spurious detections.