Digital holographic tomography (DHT) is an advanced phase-imaging-based measurement technique widely used for 3D reconstruction. However, the generated phase images often suffer from significant noise interference and irregular distortions, posing challenges for accurate reconstruction. Phase unwrapping, a critical preprocessing step for 3D reconstruction in holographic tomography, is essential to correct phase discontinuities. Traditional phase unwrapping methods frequently lack the robustness and reliability required for practical applications. To address these limitations, we explore deep learning approaches and identify that existing frameworks predominantly rely on single-stage methods, which suffer from inadequate multi-scale feature fusion and a lack of phase continuity constraints, hindering high-precision cross-scale phase unwrapping. To overcome these challenges, we propose MSSPUNet, a multi-scale, multi-stage transformer network, which leverages latent features across different scales to enhance cross-scale feature fusion. This approach achieves synergistic optimization in noise suppression, phase jump correction, and detail preservation. The network was trained on extensive simulated datasets and benchmarked against several existing phase unwrapping methods. Furthermore, we validated its performance using real DHT images of cells, organoids, phantoms, and conventional 3D-printed structures. Experimental results demonstrate that MSSPUNet offers superior accuracy, enhanced robustness, and stronger generalization capabilities compared to existing methods.
Digital holographic interferometry is a high-precision technique for quantitative imaging and measurement. However, the phase retrieved from interference patterns requires correction to obtain the true object phase. Existing numerical methods typically rely on either assumed prior constraints (as in traditional optimization-based approaches) or accurate correction data labels (as in deep learning-based methods), making it difficult to achieve an optimal balance between improving accuracy and reducing dependence on labelled data. To this end, we propose a joint neural network framework comprising a mask neural network and a phase correction neural network (MNN-PCNN) for joint background segmentation and aberration fitting to achieve phase correction in digital holographic interferometry. Inspired by the masked least-squares polynomial fitting, MNN-PCNN first employs a segmentation-based MNN to separate the pure background regions, thereby eliminating interference from the object. Subsequently, PCNN with embedded Zernike polynomial modes is leveraged to achieve accurate and reliable aberration compensation and noise suppression via self-supervised learning. Both simulation and experimental results demonstrate that the proposed MNN-PCNN outperforms classical polynomial fitting methods as well as the representative UNet. Furthermore, the developed network effectively eliminates phase deviations retrieved across multiple projection angles in digital holographic micro-tomography, which plays a critical role in enhancing the accuracy of quantitative three-dimensional measurements.