LOW-DOSE COMPUTED TOMOGRAPHY IMAGES AND CORRESPONDING DATA

国家肺筛查试验（NLST）表明，与胸部X线片相比，低剂量计算机断层扫描（LDCT）筛查受试者的肺癌死亡率大幅降低;然而，LDCT筛查也有非常高的假阳性率（FPR）。前两轮筛选的FPR约为25%，最后一轮为16%。 除了高FPR之外，还需要改进LDCT筛查阳性患者的风险预测。在NLST中，在LDCT筛查阳性并进行肺活检的患者中，约40%没有癌症。相反，还有诊断不确定性的问题，导致那些确实患有肺癌的人推迟进行活检。在21%的NLST受试者中，回顾性确定在基线LDCT扫描时存在肺癌，诊断癌症需要18个月以上。 因此，对于筛查阳性的患者中哪些人需要进行活检，以及何时进行活检的评估，还有很大的改进空间。 LDCT肺癌筛查的高FPR，沿着预测风险水平的能力有限，具有以下三个主要不利影响：1）它对接受筛查的患者造成了严重伤害，包括短期焦虑、后续CT的辐射增加以及侵入性诊断程序并发症的可能性，2）它导致医疗保健成本增加和稀缺医疗保健资源的利用增加，以及3）由于假阳性对患者和医疗保健提供者的感知和真实的负担，它用于降低LDCT筛查的采用。因此，降低FPR应有助于改善这些不利影响。 人工智能（AI）有望改变医学成像。在过去的十年中，计算机辅助检测（CAD）在帮助癌症检测和诊断方面取得了重大进展，产生了许多FDA批准的软件工具。最近，人们的努力集中在深度学习上，以开发更准确和集成的工具，这些工具可以复制或超越医疗专业人员。预计AI可以大幅降低LDCT筛查的FPR，同时最小限度地影响测试灵敏度，从而降低诊断不确定性。

The National Lung Screening Trial (NLST) demonstrated a substantial reduction in lung cancer mortality in subjects screened with low-dose computerized tomography (LDCT) as compared to chest radiographs; however, there was also a very high false positive rate (FPR) with the LDCT screens. The FPR was around 25% for the first two screening rounds, and 16% in the final round. In addition to the high FPR, there is a need for improvement in predicting risk among those with positive LDCT screens. In the NLST, of those with positive LDCT screens who went on to lung biopsy, about 40% did not have cancer. Conversely, there is also the problem of diagnostic uncertainty leading to delay in proceeding to biopsy among those who do have lung cancer. Among 21% of the NLST subjects who were retrospectively determined to have had lung cancer present at the baseline LDCT scan, it took over 18 months to diagnose the cancer. Therefore, the assessment of whom among those with positive screens needs to proceed to biopsy, and when, has room for major improvement. The high FPR of LDCT lung cancer screening, along with the limited ability in predicting risk levels, has three major detrimental effects as follows: 1) it constitutes a significant harm to patients undergoing screening in terms of short-term anxiety, increased radiation from follow-up CTs, and the potential for complications from invasive diagnostic procedures, 2) it contributes to increased health care costs and increased utilization of scarce health-care resources, and 3) it serves to lower the uptake of LDCT screening due to the perceived, and real, burden of false positives on patients and health care providers. Therefore, decreasing the FPR should serve to ameliorate these detrimental effects. Artificial intelligence (AI) is poised to transform medical imaging. In the past decade, significant progress has been made in computer aided detection (CAD) to assist with cancer detection and diagnosis, leading to a number of FDA-approved software tools. More recently, efforts are focused on deep learning to develop more accurate and integrated tools that can replicate or out-perform medical professionals. It is anticipated that AI can substantially reduce the FPR of LDCT screening while minimally affecting test sensitivity, thereby reducing diagnostic uncertainty.