Real-time AI-enhanced Low Dose Fluoroscopy

项目摘要 动机：荧光透视是5000万患者中图像引导干预不可或缺的工具 每年在美国进行的手术。它利用x射线的电离辐射， 为外科医生提供实时、高质量的成像反馈。辐射对两者都有害 病人和外科医生。重复的患者暴露已被证明会增加一倍的风险， 女性乳腺癌;同时，外科医生的暴露也令人担忧。骨科 外科医生一生中患癌症的可能性是普通外科医生的5倍;脊柱外科医生一生中患癌症的可能性是普通外科医生的12倍 与其他整形外科医生相比，辐射暴露更多。微创 近年来，外科手术被广泛采用，荧光透视的使用大大增加。 外科医生在这些手术中依赖于荧光透视提供的导航， 不能直接看到解剖结构。因此，减少辐射暴露， 荧光透视同时保持高成像质量是高度优先的。在过去几年中， 人工智能（AI）方法已经显示出有希望的进步，以提高质量， 医学成像微妙医疗公司已经获得了FDA的批准， 软件产品将PET剂量降低四倍，并提高MRI图像质量。 该提案的动机是将我们最初的成功转化为另一种成像方式， 实现低剂量透视。 方法：SBIR项目第一阶段有三个目标。目标1是使用 循环的深度学习架构，以实现荧光透视的6倍剂量减少。目标二是 设计模型修剪、内核优化和高性能推理框架， 实现实时处理。最后，在目标3中，我们将定性评估和 对所开发的软件在体模和尸体上进行了定量分析。 意义：这项工作将使荧光透视剂量降低六倍。的完成 该项目将产生广泛的影响，大大减少手术室的辐射暴露，因此 降低患者和临床医生患癌症的风险。

Project Summary Motivation: Fluoroscopy is an indispensable tool for image-guided interventions in 50 million surgeries performed annually in the US. It leverages ionizing radiation from x-rays to provide surgeons with real-time, high-quality imaging feedback. Radiation exposure is harmful for both the patients and surgeons. Repetitive patient exposure has been shown to double the risk of breast cancer in women; meanwhile, surgeon exposure is also concerning. Orthopedic surgeons are 5x more likely to develop cancer in their lifetime; spine surgeons receive up to 12x more radiation exposure compared to other orthopedic surgeons. With minimally invasive surgery becoming widely adopted in recent years, the use of fluoroscopy has greatly increased. Surgeons rely on the navigation provided by fluoroscopy during these procedures, as they do not have direct visualization of the anatomy. Therefore, reducing the radiation exposure from fluoroscopy while maintaining high imaging quality is a high priority. In the past few years, artificial intelligence (AI) methods have shown promising advances to improve the quality of medical imaging. Subtle Medical Inc. has already received FDA clearance for its AI-based software products to reduce the dose for PET by four times and improve image quality for MRI. The motivation of this proposal is to translate our initial success to another imaging modality and achieve low dose fluoroscopy. Approach: This phase I SBIR project has three aims. Aim 1 is to develop AI software using recurrent deep learning architecture to achieve 6x dose reduction for fluoroscopy. Aim 2 is to design model pruning, kernel optimization, and high-performance inference frameworks to achieve real-time processing. Finally, in Aim 3, we will evaluate both qualitatively and quantitatively the developed software on phantoms and cadavers. Significance: This work will enable six times lower dose fluoroscopy. The completion of the project will have wide impact to greatly reduce radiation exposure in the operating room, hence reducing the risk of cancer development for both patients and clinicians.