Background

Although deep learning has brought revolutionary changes in health care, reliance on manually selected cross-sectional images and segmentation remain methodological barriers.

Purpose

To develop and validate an automated preoperative artificial intelligence (AI) algorithm for tumor and lymph node (LN) segmentation with CT imaging for prediction of LN metastasis in patients with pancreatic ductal adenocarcinoma (PDAC).

Materials and Methods

In this retrospective study, patients with surgically resected, pathologically confirmed PDAC underwent multidetector CT from January 2015 to April 2020. Three models were developed, including an AI model, a clinical model, and a radiomics model. CT-determined LN metastasis was diagnosed by radiologists. Multivariable logistic regression analysis was conducted to develop the clinical and radiomics models. The performance of the models was determined on the basis of their discrimination and clinical utility. Kaplan-Meier curves, the log-rank test, or Cox regression were used for survival analysis.

Results

Overall, 734 patients (mean age, 62 years ± 9 [SD]; 453 men) were evaluated. All patients were split into training (n = 545) and validation (n = 189) sets. Patients who had LN metastasis (LN-positive group) accounted for 340 of 734 (46%) patients. In the training set, the AI model showed the highest performance (area under the receiver operating characteristic curve [AUC], 0.91) in the prediction of LN metastasis, whereas the radiologists and the clinical and radiomics models had AUCs of 0.58, 0.76, and 0.71, respectively. In the validation set, the AI model showed the highest performance (AUC, 0.92) in the prediction of LN metastasis, whereas the radiologists and the clinical and radiomics models had AUCs of 0.65, 0.77, and 0.68, respectively (P < .001). AI model–predicted positive LN metastasis was associated with worse survival (hazard ratio, 1.46; 95% CI: 1.13, 1.89; P = .004).

Conclusion

An artificial intelligence model outperformed radiologists and clinical and radiomics models for prediction of lymph node metastasis at CT in patients with pancreatic ductal adenocarcinoma.

© RSNA, 2022

Online supplemental material is available for this article.

See also the editorial by Chu and Fishman in this issue.

中文翻译：

---

人工智能预测胰腺导管腺癌 CT 淋巴结转移

 背景

尽管深度学习给医疗保健带来了革命性的变化，但对手动选择横截面图像和分割的依赖仍然存在方法论障碍。

 目的

开发和验证一种自动术前人工智能 (AI) 算法，用于通过 CT 成像进行肿瘤和淋巴结 (LN) 分割，以预测胰腺导管腺癌 (PDAC) 患者的 LN 转移。

 材料和方法

在这项回顾性研究中，2015年1月至2020年4月，手术切除并经病理证实的PDAC患者接受了多排CT检查。开发了三种模型，包括AI模型、临床模型和放射组学模型。 CT 确定的淋巴结转移由放射科医生诊断。进行多变量逻辑回归分析以开发临床和放射组学模型。模型的性能是根据其区分度和临床实用性来确定的。 Kaplan-Meier 曲线、对数秩检验或 Cox 回归用于生存分析。

 结果

总体而言，对 734 名患者（平均年龄，62 岁 ± 9 [SD]；453 名男性）进行了评估。所有患者均分为训练组 ( n = 545) 和验证组 ( n = 189)。 734 例患者中，有 LN 转移的患者（LN 阳性组）占 340 例（46%）。在训练集中，AI 模型在预测 LN 转移方面表现出最高的性能（受试者工作特征曲线下面积 [AUC] 为 0.91），而放射科医生以及临床和放射组学模型的 AUC 分别为 0.58、0.76 和分别为 0.71。在验证集中，AI 模型在预测 LN 转移方面表现出最高的性能（AUC，0.92），而放射科医生以及临床和放射组学模型的 AUC 分别为 0.65、0.77 和 0.68（ P < .001） 。 AI 模型预测的阳性淋巴结转移与较差的生存率相关（风险比，1.46；95% CI：1.13，1.89； P = .004）。

 结论

在预测胰腺导管腺癌患者的 CT 淋巴结转移方面，人工智能模型的表现优于放射科医生以及临床和放射组学模型。

 © 北美放射学会，2022

本文提供在线补充材料。

另请参阅本期 Chu 和 Fishman 的社论。