Scholarlink

<p>Delivering optoacoustic (OA) imaging to point‑of‑care and resource‑limited settings requires systems that are both compact and affordable without compromising image quality. Conventional OA design treats hardware configuration and image reconstruction as separate, iterative tasks, often resulting in redundant components and suboptimal performance. Here we introduce a deep learning‑based codesign framework that unifies these processes through a differentiable forward model, enabling joint optimization of transducer layout and reconstruction algorithms. This approach yields ultrasparse circular and hemispherical arrays that use only 0.8–12.5% of conventional detector counts while maintaining state‑of‑the‑art spatial and temporal fidelity. Guided by the learned designs, we built a 32‑element handheld OA system powered by low‑cost laser diodes, achieving ultrafast volumetric imaging at 2 kHz. By integrating artificial intelligence‑driven optimization directly into system architecture, this work establishes a pathway towards compact, economical OA instruments with strong translational potential.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Deep learning codesign for ultrasparse, high‑performance optoacoustic imaging

  • 期刊论文

  • 发表日期：

    2026年5月7日

查看全文
  • Xuyang Chang,
  • Xiang Liu,
  • Federico Villani,
  • Quanyu Zhou,
  • Tian Jin,
  • Lin Tang,
  • Luca Benini,
  • Xosé Luís Deán-Ben,
  • Liheng Bian,
  • Zhenyue Chen,
  • Daniel Razansky

摘要

Delivering optoacoustic (OA) imaging to point‑of‑care and resource‑limited settings requires systems that are both compact and affordable without compromising image quality. Conventional OA design treats hardware configuration and image reconstruction as separate, iterative tasks, often resulting in redundant components and suboptimal performance. Here we introduce a deep learning‑based codesign framework that unifies these processes through a differentiable forward model, enabling joint optimization of transducer layout and reconstruction algorithms. This approach yields ultrasparse circular and hemispherical arrays that use only 0.8–12.5% of conventional detector counts while maintaining state‑of‑the‑art spatial and temporal fidelity. Guided by the learned designs, we built a 32‑element handheld OA system powered by low‑cost laser diodes, achieving ultrafast volumetric imaging at 2 kHz. By integrating artificial intelligence‑driven optimization directly into system architecture, this work establishes a pathway towards compact, economical OA instruments with strong translational potential.