Departments of Radiology, University of Wisconsin-Madison, Madison, Wisconsin.
Departments of Radiology, University of Wisconsin-Madison, Madison, Wisconsin; Urology, University of Wisconsin-Madison, Madison, Wisconsin.
J Vasc Interv Radiol. 2024 May;35(5):770-779.e1. doi: 10.1016/j.jvir.2023.11.016. Epub 2023 Nov 24.
To evaluate the concordance between lung biopsy puncture pathways determined by artificial intelligence (AI) and those determined by expert physicians.
An AI algorithm was created to choose optimal lung biopsy pathways based on segmented thoracic anatomy and emphysema in volumetric lung computed tomography (CT) scans combined with rules derived from the medical literature. The algorithm was validated using pathways generated from CT scans of randomly selected patients (n = 48) who had received percutaneous lung biopsies and had noncontrast CT scans of 1.25-mm thickness available in picture archiving and communication system (PACS) (n = 28, mean age, 68.4 years ± 9.2; 12 women, 16 men). The algorithm generated 5 potential pathways per scan, including the computer-selected best pathway and 4 random pathways (n = 140). Four experienced physicians rated each pathway on a 1-5 scale, where scores of 1-3 were considered safe and 4-5 were considered unsafe. Concordance between computer and physician ratings was assessed using Cohen's κ.
The algorithm ratings were statistically equivalent to the physician ratings (safe vs unsafe: κ¯=0.73; ordinal scale: κ¯=0.62). The computer and physician ratings were identical in 57.9% (81/140) of cases and differed by a median of 0 points. All least-cost "best" pathways generated by the algorithm were considered safe by both computer and physicians (28/28) and were judged by physicians to be ideal or near ideal.
AI-generated lung biopsy puncture paths were concordant with expert physician reviewers and considered safe. A prospective comparison between computer- and physician-selected puncture paths appears indicated in addition to expansion to other anatomic locations and procedures.
评估人工智能（AI）确定的肺活检穿刺路径与专家医生确定的路径之间的一致性。
创建了一个 AI 算法，根据分段的胸部解剖结构和体积 CT 扫描中的肺气肿，以及从医学文献中得出的规则，选择最佳的肺活检途径。该算法使用从接受经皮肺活检的随机选择的患者的 CT 扫描中生成的途径进行验证（n=48），并且在 PACS 中有可用的非对比 1.25-mm 厚度 CT 扫描（n=28，平均年龄 68.4 岁±9.2；12 名女性，16 名男性）。该算法为每个扫描生成了 5 条潜在的途径，包括计算机选择的最佳途径和 4 条随机途径（n=140）。4 名经验丰富的医生对每条途径进行了 1-5 分的评分，其中 1-3 分被认为是安全的，4-5 分被认为是不安全的。使用 Cohen's κ 评估计算机和医生评分之间的一致性。
算法评分与医生评分统计学等效（安全与不安全：κ¯=0.73；有序量表：κ¯=0.62）。在 57.9%（81/140）的情况下，计算机和医生的评分是相同的，差异中位数为 0 分。算法生成的所有最低成本“最佳”途径都被计算机和医生认为是安全的（28/28），并且被医生判断为理想或接近理想。
AI 生成的肺活检穿刺路径与专家医生评审者一致，并被认为是安全的。除了扩展到其他解剖位置和程序外，还需要在计算机和医生选择的穿刺路径之间进行前瞻性比较。
