Barnes Amy, White Rebecca, Venables Heather, Lam Vincent, Vaidhyanath Ram
Consultant Radiographer, University Hospitals of Leicester NHS Trust, Leicester, UK.
Senior Lecturer, University of Derby, Derby, UK.
Ultrasound. 2024 Dec 7:1742271X241299220. doi: 10.1177/1742271X241299220.
This pilot study aims to evaluate the clinical impact of artificial intelligence-based decision support, Koios Decision Support™, on the diagnostic performance of ultrasound assessment of thyroid nodules, and as a result to avoid fine needle aspiration.
This retrospective pilot study was conducted on ultrasound images of thyroid nodules investigated with fine needle aspiration from January 2022 to December 2022. Orthogonal ultrasound images of thyroid nodules, previously investigated with fine needle aspiration, were compared with the Koios Decision Support™ suggestion to perform fine needle aspiration. Surgical histology was used as ground truth.
A total of 29 patients (76% women) with a mean age of 48 ± 16.5 years were evaluated, = 15 (52%) were histologically proven benign and = 14 (48%) were malignant. In the benign group, Koios Decision Support™ suggested avoidable fine needle aspiration in = 8 (53%). In the malignant group, Koios Decision Support™ suggested follow-up or no fine needle aspiration in = 2 (14%). Sensitivity is 85.7% ( = 12) ( = 0.027), whereas specificity is 53.3% ( = 8) ( = 0.027). The positive predictive value is 63.2% ( = 12), negative predictive value is 80% ( = 8), false-negative value is 20% ( = 2) and false-positive value is 36.8% ( = 7). Based on artificial intelligence decision, one cancer would have been missed.
Artificial intelligence can improve specificity without significantly compromising sensitivity. There was a suggested reduction in the fine needle aspiration rate, in the histologically proven benign nodules, by 53%. This had no statistical significance, likely due to the small population, however, it is thought to be the largest study to date. Further investigation with wider-ranging studies is suggested.
本初步研究旨在评估基于人工智能的决策支持系统Koios Decision Support™对甲状腺结节超声评估诊断性能的临床影响，从而避免细针穿刺活检。
本回顾性初步研究针对2022年1月至2022年12月期间接受细针穿刺活检的甲状腺结节超声图像进行。将先前接受过细针穿刺活检的甲状腺结节的正交超声图像与Koios Decision Support™关于进行细针穿刺活检的建议进行比较。手术组织学结果作为金标准。
共评估了29例患者（76%为女性），平均年龄48±16.5岁，其中15例（52%）经组织学证实为良性，14例（48%）为恶性。在良性组中，Koios Decision Support™建议进行了8例（53%）可避免的细针穿刺活检。在恶性组中，Koios Decision Support™建议进行随访或不进行细针穿刺活检的有2例（14%）。敏感性为85.7%（12例）（P = 0.027），而特异性为53.3%（8例）（P = 0.027）。阳性预测值为63.2%（12例），阴性预测值为80%（8例），假阴性值为20%（2例），假阳性值为36.8%（7例）。基于人工智能决策，有1例癌症可能被漏诊。
人工智能可以提高特异性，而不会显著降低敏感性。在组织学证实为良性的结节中，细针穿刺活检率建议降低53%。这没有统计学意义，可能是由于样本量小，但这被认为是迄今为止规模最大的研究。建议进行更广泛的进一步研究。
