Unrolled deep learning for breast cancer detection using limited-view photoacoustic tomography data

Photoacoustic tomography (PAT) has emerged as a promising imaging modality for breast cancer detection, offering unique advantages in visualizing tissue composition without ionizing radiation. However, limited-view scenarios in clinical settings present significant challenges for image reconstruction quality and computational efficiency. This paper introduces novel unrolled deep learning networks based on split Bregman total variation (SBTV) and relaxed basis pursuit alternating direction method of multipliers (rBP-ADMM) algorithms to address these challenges. Our approach combines transfer learning from full-view to limited-view scenarios with U-Net denoiser integration, achieving state-of-the-art reconstruction quality (MS-SSIM> 0.95) while reducing reconstruction time by 92% compared to traditional methods. The effectiveness of different sensor configurations is analyzed through restricted isometry property (RIP) analysis and coherence values, demonstrating that semicircular arrays achieve a RIP constant of 0.76 and coherence of 0.77, closely approximating full-view performance (RIP: 0.75, coherence: 0.78). These metrics validate the theoretical foundation for accurate sparse signal recovery in limited-view scenarios. Comprehensive evaluations across semicircular, concave, and convex sensor arrangements show that the proposed U-SBTV network consistently outperforms existing methods, particularly when combined with the U-Net denoiser. This advancement in limited-view PAT reconstruction brings the technology closer to practical clinical application, potentially improving early breast cancer detection capabilities.