Deep Learning-Enabled Cross-Device Calibration for Smartphone-Based Optical Biosensing: A Review

Smartphone-based optical biosensing has become a compelling route toward portable, low-cost, and accessible analysis, yet its broader translation is still constrained by a central unresolved bottleneck: poor cross-device analytical consistency. Variations in illumination, optics, sensor spectral responsivity, image signal processing (ISP), acquisition geometry, and assay dynamics can transform the same biochemical event into substantially different digital readouts, thereby undermining quantitative reliability. This review recasts deep learning-enabled cross-device calibration as an optical readout harmonization problem for transferable quantification rather than a generic prediction task. We synthesize the field through a unified three-level framework comprising observation calibration to standardize device-dependent images, feature calibration to learn assay-relevant and device-robust representations, and model calibration to impose physics-aware and uncertainty-aware constraints on interpretation. Across smartphone optical biosensing, deep learning shows clear promise, but the evidence base remains uneven, and deployment is still limited less by the absence of models than by weak benchmarking and incomplete validation under device, illumination, lot/time, and field-use shifts. By defining the evidentiary requirements for deployment and providing a practical framework for benchmark design, challenge-oriented evaluation, and revalidation, this review offers a sharper foundation for developing reliable, scalable, and analytically defensible smartphone biosensing systems.