Joint signal-domain segmentation and reconstruction for photoacoustic tomography via physics-informed self-supervised learning

Abstract

Objective. Conventional photoacoustic tomography (PAT) follows a decoupled pipeline of image reconstruction and segmentation, causing error propagation and information loss as artifacts and reduced contrast in reconstructed images degrade subsequent segmentation accuracy. Approach. We propose the Joint Image Reconstruction and Segmentation in the Signal Domain Network (JIRSSD-Net), an end-to-end framework that unifies segmentation and reconstruction by performing feature extraction directly on raw acoustic pressure signals prior to image formation. JIRSSD-Net employs a dynamic physics-guided self-supervision mechanism that progressively relaxes time-of-flight (TOF) physical priors by annealing their weighting coefficient from 1.0 to 0.2 via a cosine schedule as data-driven representations mature. This mechanism comprises three coupled components: a physics-driven module estimating initial target boundaries through TOF analysis providing an early geometric anchor; a self-supervised module with dual-network contrastive learning, where physical priors serve as a progressively annealed curriculum; and a cross-domain gradient pathway from reconstruction loss back to the signal-domain feature extractor. A reconstruction subnetwork fuses segmented signal features with an initial low-quality image to produce a high-fidelity output. Main results. JIRSSD-Net was validated on simulated, phantom, and in vivo murine datasets, outperforming image-domain, signal-domain, and joint baselines across all metrics with statistical significance confirmed by the Friedman test. On in vivo data, it surpassed the state-of-the-art joint method BFIO-Net in both reconstruction and segmentation, achieving approximately 3.6% and 3.4% higher structural similarity index and peak signal-to-noise ratio (PSNR), while improving the Jaccard index by 7.9% and reducing the average symmetric surface distance by 41.3%. JIRSSD-Net also demonstrated superior robustness under realistic noise with high computational efficiency. Significance. By coupling physical priors with self-supervised learning through a dynamic annealing mechanism, JIRSSD-Net eliminates the error propagation of conventional sequential pipelines and enhances robustness and accuracy for preclinical PAT applications.