Multitarget stool DNA(mt-sDNA) testing was approved for average risk colorectal cancer(CRC) screening by the United States Food and Drug Administration and thereafter reimbursed for use by the Medicare program(2014).T...Multitarget stool DNA(mt-sDNA) testing was approved for average risk colorectal cancer(CRC) screening by the United States Food and Drug Administration and thereafter reimbursed for use by the Medicare program(2014).The United States Preventive Services Task Force(USPSTF) October 2015 draft recommendation for CRC screening included mt-s DNA as an "alternative" screening test that "may be useful in select clinical circumstances",despite its very high sensitivity for early stage CRC.The evidence supporting mt-s DNA for routine screening use is robust.The clinical efficacy of mt-s DNA as measured by sensitivity,specificity,life-years gained(LYG),and CRC deaths averted is similar to or exceeds that of the other more specifically recommended screening options included in the draft document,especially those requiring annual testing adherence.In a population with primarily irregular screening participation,tests with the highest point sensitivity and reasonable specificity are more likely to favorably impact CRC related morbidity and mortality than those depending on annual adherence.This paper reviews the evidence supporting mt-s DNA for routine screening and demonstrates,using USPSTF's modeling data,that mt-s DNA at three-year intervals provides significant clinical net benefits and fewer complications per LYG than annual fecal immunochemical testing,high sensitivity guaiac based fecal occult blood testing and 10-year colonoscopy screening.展开更多
芯粒集成逐渐成为不同场景下敏捷定制深度学习芯片的高可扩展性的解决方案,芯片设计者可以通过集成设计、验证完成的第三方芯粒来降低芯片开发周期和成本,提高芯片设计的灵活性和芯片良率.在传统的芯片设计和商业模式中,编译器等专用软...芯粒集成逐渐成为不同场景下敏捷定制深度学习芯片的高可扩展性的解决方案,芯片设计者可以通过集成设计、验证完成的第三方芯粒来降低芯片开发周期和成本,提高芯片设计的灵活性和芯片良率.在传统的芯片设计和商业模式中,编译器等专用软件工具链是芯片解决方案的组成部分,并在芯片性能和开发中发挥重要作用.然而,当使用第三方芯粒进行芯片敏捷定制时,第三方芯粒所提供的专用工具链无法预知整个芯片的资源,因此无法解决敏捷定制的深度学习芯片的任务部署问题,而为敏捷定制的芯片设计全新的工具链需要大量的时间成本,失去了芯片敏捷定制的优势.因此,提出一种面向深度学习集成芯片的可扩展框架(scalable framework for integrated deep learning chips)--Puzzle,它包含从处理任务输入到运行时管理芯片资源的完整流程,并自适应地生成高效的任务调度和资源分配方案,降低冗余访存和芯粒间通信开销.实验结果表明,该可扩展框架为深度学习集成芯片生成的任务部署方案可自适应于不同的工作负载和硬件资源配置,与现有方法相比平均降低27.5%的工作负载运行延迟.展开更多
文摘Multitarget stool DNA(mt-sDNA) testing was approved for average risk colorectal cancer(CRC) screening by the United States Food and Drug Administration and thereafter reimbursed for use by the Medicare program(2014).The United States Preventive Services Task Force(USPSTF) October 2015 draft recommendation for CRC screening included mt-s DNA as an "alternative" screening test that "may be useful in select clinical circumstances",despite its very high sensitivity for early stage CRC.The evidence supporting mt-s DNA for routine screening use is robust.The clinical efficacy of mt-s DNA as measured by sensitivity,specificity,life-years gained(LYG),and CRC deaths averted is similar to or exceeds that of the other more specifically recommended screening options included in the draft document,especially those requiring annual testing adherence.In a population with primarily irregular screening participation,tests with the highest point sensitivity and reasonable specificity are more likely to favorably impact CRC related morbidity and mortality than those depending on annual adherence.This paper reviews the evidence supporting mt-s DNA for routine screening and demonstrates,using USPSTF's modeling data,that mt-s DNA at three-year intervals provides significant clinical net benefits and fewer complications per LYG than annual fecal immunochemical testing,high sensitivity guaiac based fecal occult blood testing and 10-year colonoscopy screening.
文摘芯粒集成逐渐成为不同场景下敏捷定制深度学习芯片的高可扩展性的解决方案,芯片设计者可以通过集成设计、验证完成的第三方芯粒来降低芯片开发周期和成本,提高芯片设计的灵活性和芯片良率.在传统的芯片设计和商业模式中,编译器等专用软件工具链是芯片解决方案的组成部分,并在芯片性能和开发中发挥重要作用.然而,当使用第三方芯粒进行芯片敏捷定制时,第三方芯粒所提供的专用工具链无法预知整个芯片的资源,因此无法解决敏捷定制的深度学习芯片的任务部署问题,而为敏捷定制的芯片设计全新的工具链需要大量的时间成本,失去了芯片敏捷定制的优势.因此,提出一种面向深度学习集成芯片的可扩展框架(scalable framework for integrated deep learning chips)--Puzzle,它包含从处理任务输入到运行时管理芯片资源的完整流程,并自适应地生成高效的任务调度和资源分配方案,降低冗余访存和芯粒间通信开销.实验结果表明,该可扩展框架为深度学习集成芯片生成的任务部署方案可自适应于不同的工作负载和硬件资源配置,与现有方法相比平均降低27.5%的工作负载运行延迟.
