Deep Learning-Based Phase Unwrapping and Image Denoising for 3D Reconstruction in Digital Holography and Photoacoustic Imaging

Abstract

Full-field optical and photoacoustic imaging systems—such as Digital Holography and Full-Field Photoacoustic Tomography (FF-PAT)—have become essential tools for non-invasive, high-resolution imaging in biomedical diagnostics and industrial inspection. These systems capture critical phase and displacement information necessary for 3D reconstruction and quantitative analysis. However, phase and displacement maps are often severely corrupted by speckle noise, Gaussian noise, and ambient interference, posing significant challenges for accurate signal recovery. Traditional denoising and unwrapping algorithms often fail under these harsh conditions due to their sensitivity to noise, assumptions about phase smoothness, and high computational costs. This thesis proposes three deep learning-based frameworks to address these challenges, offering robust, accurate, and efficient solutions for denoising and phase recovery in full-field imaging. Each framework is specifically tailored to overcome the limitations of conventional methods and to operate reliably under low- SNR conditions. First, we propose WPD-Net (Wrapped Phase Denoising Network), a lightweight neural network designed for denoising wrapped-phase images affected by complex noise mixtures. WPD-Net integrates Residual Dense Attention Blocks (RDABs) to selectively enhance important features while suppressing irrelevant noise. A growth-rate-based multi-scale feature expansion and dense feature fusion strategy are incorporated to preserve fine structural details and ensure phase continuity. Evaluations on both synthetic and experimental datasets demonstrate that WPD-Net outperforms traditional and deep learning- based denoising approaches in terms of PSNR, SSIM, and visual quality. Its compact design also enables real- time inference, making it highly suitable for biomedical and industrial optical systems where fast, accurate processing is critical. Next, to unify denoising and phase unwrapping into a single step, we introduce DenSFA- PU (Densely Connected Spatial Feature Aggregator for Phase Unwrapping). DenSFA-PU is an end-to-end regression model that directly maps noisy wrapped-phase inputs to unwrapped continuous outputs. The network combines dense connectivity with a Spatial Feature Aggregator (SFA) module, integrating Bi- directional LSTM layers and Bottleneck Attention Modules (BAMs) to capture long-range dependencies and focus on structurally important regions. Extensive experiments show that DenSFA-PU achieves superior results across standard evaluation metrics (PSNR, SSIM, NRMSE) and maintains fast inference times (~29.31 ms per image). This efficiency, combined with high robustness against severe noise, positions DenSFA-PU as a powerful tool for real-time and high-throughput phase imaging applications. Finally, for FF-PAT, we address displacement map denoising—a particularly challenging problem due to the extremely low SNR of photoacoustic signals. We propose an attention-guided, multi-scale feature fusion network that combines RDABs with dual channel and spatial attention modules. Trained exclusively on experimentally acquired FF- PAT data, the model demonstrates strong generalization to real-world noise conditions. It achieves a PSNR of 34.25 dB, significantly outperforming coherent averaging (13.88 dB) and U-Net-based approaches (26.87 dB). With a processing speed of 0.53 seconds per displacement map, the model enables real-time volumetric photoacoustic imaging while preserving critical structural information. ⓒ2025 Muhammad Awais ALL RIGHTS RESERVED