Complex-valued convolutional neural networks for disease detection utilizing digital holographic wavefronts

Abstract Phase carries crucial morphological and compositional information and is central to many optical imaging modalities. Digital holographic microscopy (DHM), interferometric technique, captures the complex object field of a specimen, containing both phase and amplitude. Conventional deep learning networks process complex data as paired real channels, neglecting intrinsic phase–amplitude coupling. In this reported work, we introduce complex-valued convolutional neural networks (CV-CNNs) for binary classification of normal and cancerous cervical cells using the DHM complex object field. Four CV-CNN variants (ModReLU, CReLU, Cardioid, zReLU) were evaluated via five-fold cross-validation against an equivalent real-valued CNN. The ModReLU CV-CNN achieved the highest performance (Accuracy = 0.89 ± 0.03, F1 = 0.89 ± 0.04, Precision = 0.89 ± 0.04, Recall = 0.89 ± 0.05, AUC = 0.96 ± 0.02), demonstrating that phase-aware complex convolution significantly enhances the classification performance. By eliminating phase unwrapping, it also reduces associated computational cost and ambiguity. The methodology is also applicable to other phase-sensitive imaging modalities.