Self-supervised photoacoustic tomography image reconstruction based on multi-curve forward projection and implicit neural representation

In sparse-view or limited-view photoacoustic tomography (PAT), the imaging performance is mainly limited by image reconstruction accuracy. Integrating PAT’s multi-curve forward model with implicit neural representation (INR) presents a promising approach to significantly enhance PAT imaging performance. Inspired by the recently reported INR concept, we propose a self-supervised image reconstruction method that combines the multi-curve forward model of PAT with INR, which we named MIN-PAT. This method aims to improve reconstruction quality using only undersampled data. The proposed MIN-PAT framework extracts spatial coordinates from geometric projection curves in the forward model and employs INR to learn an implicit mapping from image coordinates to pixel intensity, eliminating the need for fully sampled ground-truth data. Experimental results under various sparse-view and limited-view conditions demonstrate that MIN-PAT effectively recovers target structures while suppressing artifacts compared to conventional reconstruction methods. By synergizing PAT’s multi-curve forward model with INR-based learning, MIN-PAT enables robust, self-supervised reconstruction that enhances image quality without external training data. This approach shows significant potential for improving PAT imaging in challenging acquisition scenarios. Code available at: https://github.com/CbinHu/MIN-PAT . • MIN-PAT: a self-supervised method for limited-view and sparse-view PAT reconstruction without fully sampled data. • MIN-PAT fuses multi-curve forward model with implicit neural representation, enhancing quality and reducing artifacts in undersampled PAT. • Simulated and in vivo experiments demonstrate MIN-PAT’s superior performance, robustness, and practical potential for PAT imaging.