3D Segment Anything Model for 3D Protoacoustic Imaging Enhancement and Dose Verification in Proton Therapy

Protoacoustic (PA) imaging captures acoustic signals generated during proton interactions and shows strong potential for real-time monitoring of dose distribution, addressing the challenges in proton therapy where accuracy can be affected by patient positioning and anatomical changes. This study investigates the use of SAM-Med3D, a large-scale 3D segmentation model, to enhance PA image reconstruction and address the limited-angle problem that impairs spatial resolution. We modified the SAM-Med3D architecture by introducing a global-local feature fusion strategy in the encoder to preserve structural details and replacing the prompt-based decoder with a lightweight design that progressively restores image resolution. The method was evaluated using a dataset of prostate cancer patients, incorporating CT images and clinical treatment plans, as well as an experimental pencil beam dataset in a water tank. Our model achieved significant improvements over baseline methods, with average RMSE and SSIM values of 0.017 and 0.958 for image reconstruction, and 0.013 and 0.963 for dose verification, respectively (p < 0.05). Gamma index evaluation showed high agreement (over 97%) between predicted and ground truth dose maps. The method took approximately 1 second per patient, supporting its feasibility for online 3D dose verification. Results show proposed method provides an effective solution to the limited-angle challenge in PA imaging and contributes to the development of high-precision, real-time 3D dose verification in proton therapy. It also underscores the potential of large-scale models like SAM-Med3D for advancing image enhancement techniques in medical imaging, particularly in the context of real-time radiation therapy monitoring.