3D Protoacoustic Image Reconstruction Using a SAM-Med3D Foundation Model
Abstract
Purpose
Protoacoustic (PA) imaging enables real-time verification of proton dose deposition by reconstructing 3D initial pressure maps from acoustic signals generated during proton delivery. However, in practical treatment settings, detectors cannot fully surround the patient, and only limited-angle acquisition can be achieved, leading to severe distortion and artifacts and limiting its accuracy for dose verification.
Methods
We propose a two-stage reconstruction–enhancement framework for limited-angle 3D PA imaging based on a large foundation model. Building on the Segment-Anything-Model (SAM), we adapt its 3D medical variant (SAM-Med3D) for volumetric PA reconstruction. In the reconstruction stage, a transformer-based network (ReconSAM) directly maps raw radiofrequency (RF) signals to a coarse 3D initial pressure volume using lightweight adapter modules for task-specific feature adaptation, while keeping most pretrained encoder weights frozen. In the enhancement stage, a second-stage network refines the coarse reconstruction to suppress limited-angle artifacts and restore sharper anatomical and dose-related structures. The method is evaluated on simulated prostate PA datasets from 126 anonymized patients with clinical treatment plans, with RF signals generated using k-Wave under heterogeneous tissue modeling and realistic noise. Patients’ data were split into 80/20/26 for training/validation/testing. Performance of the model was assessed for both accumulated clinical fields and individual pencil-beam cases and compared against both time-reversal and prior learning-based reconstruction methods.
Results
The proposed method consistently improves reconstruction fidelity under limited-angle acquisition. For accumulated-field PA reconstruction, it achieves 35.43 dB PSNR and 0.94 SSIM, substantially outperforming time-reversal (24.02 dB PSNR, 0.85 SSIM) and a transformer-based baseline (30.37 dB PSNR, 0.93 SSIM). For pencil-beam reconstruction, the method reaches 48.65 dB PSNR and 0.99 SSIM, demonstrating improved preservation of sharp beam boundaries critical for accurate proton range verification.
Conclusion
This foundation-model-driven framework mitigates limited-angle artifacts and enables accurate, fast 3D PA reconstruction, greatly enhancing its value for online proton dose verification