Deep Learning-Based Simultaneous Cancerous and Tuberculosis Cells Detection Biosensor: A Computational Approach

Early, accurate, and cost-effective detection of multiple diseases remains a critical challenge in modern biomedical diagnostics. This study offers a novel computational framework for the simultaneous detection of cancerous and tuberculosis cells using a gold-coated photonic crystal fiber (PCF)-based surface plasmon resonance (SPR) biosensor integrated with deep learning (DL). The proposed dual-channel biosensor structure, optimized through the Finite Element Method (FEM), is designed to detect multi-analyte samples by analyzing confinement loss (CL) across different refractive indices (RI). To enhance prediction accuracy and support rapid parameter optimization, a generative adversarial network (GAN) model was designed to estimate CL based on key sensor design features. The GAN model achieved superior performance, with a mean squared error (MSE) of 0.0175, a mean absolute error (MAE) of 0.1250, and an R2 of 0.9087, compared to traditional machine learning (ML) models, such as decision trees and random forests. Explainability was incorporated through SHAP (SHapley Additive exPlanations) analysis, which identified critical design parameters that influence model output, thus enhancing transparency and trustworthiness. Extensive ablation studies with various cancer and tuberculosis cells validated the reliability of the proposed model. The predicted CL curves closely aligned with simulated results, confirming the robustness of the proposed model in real-time dual-analyte detection. This study offers a promising data-driven strategy for designing multi-analyte biosensors, paving the way for next-generation noninvasive diagnostic tools.