基于全幻灯片图像的 T1 结直肠癌淋巴结转移的无监督人工智能预测。

Whole slide image-based prediction of lymph node metastasis in T1 colorectal cancer using unsupervised artificial intelligence.

机构信息

Digestive Disease Center, Showa University Northern Yokohama Hospital, Kanagawa, Japan.

Division of Gastroenterology and Hepatology, National University Hospital, Singapore City, Singapore.

出版信息

Dig Endosc. 2023 Nov;35(7):902-908. doi: 10.1111/den.14547. Epub 2023 Apr 10.

DOI:10.1111/den.14547

PMID:36905308

Abstract

OBJECTIVES

Lymph node metastasis (LNM) prediction for T1 colorectal cancer (CRC) is critical for determining the need for surgery after endoscopic resection because LNM occurs in 10%. We aimed to develop a novel artificial intelligence (AI) system using whole slide images (WSIs) to predict LNM.

METHODS

We conducted a retrospective single center study. To train and test the AI model, we included LNM status-confirmed T1 and T2 CRC between April 2001 and October 2021. These lesions were divided into two cohorts: training (T1 and T2) and testing (T1). WSIs were cropped into small patches and clustered by unsupervised K-means. The percentage of patches belonging to each cluster was calculated from each WSI. Each cluster's percentage, sex, and tumor location were extracted and learned using the random forest algorithm. We calculated the areas under the receiver operating characteristic curves (AUCs) to identify the LNM and the rate of over-surgery of the AI model and the guidelines.

RESULTS

The training cohort contained 217 T1 and 268 T2 CRCs, while 100 T1 cases (LNM-positivity 15%) were the test cohort. The AUC of the AI system for the test cohort was 0.74 (95% confidence interval [CI] 0.58-0.86), and 0.52 (95% CI 0.50-0.55) using the guidelines criteria (P = 0.0028). This AI model could reduce the 21% of over-surgery compared to the guidelines.

CONCLUSION

We developed a pathologist-independent predictive model for LNM in T1 CRC using WSI for determination of the need for surgery after endoscopic resection.

TRIAL REGISTRATION

UMIN Clinical Trials Registry (UMIN000046992, https://center6.umin.ac.jp/cgi-open-bin/ctr/ctr_view.cgi?recptno=R000053590).

摘要

目的

预测 T1 结直肠癌（CRC）的淋巴结转移（LNM）对于确定内镜切除术后是否需要手术至关重要，因为 10%的患者存在 LNM。我们旨在开发一种使用全切片图像（WSI）的新型人工智能（AI）系统来预测 LNM。

方法

我们进行了一项回顾性单中心研究。为了训练和测试 AI 模型，我们纳入了 2001 年 4 月至 2021 年 10 月期间经证实存在 LNM 的 T1 和 T2CRC。这些病变分为两个队列：训练（T1 和 T2）和测试（T1）。将 WSI 裁剪成小的斑块，并通过无监督 K-means 聚类。从每个 WSI 计算属于每个簇的斑块百分比。从每个簇的百分比、性别和肿瘤位置中提取并使用随机森林算法进行学习。我们计算了接收者操作特征曲线（ROC）下的面积（AUCs），以确定 AI 模型和指南的 LNM 和过度手术率。

结果

训练队列包含 217 例 T1 和 268 例 T2CRC，而 100 例 T1 病例（LNM 阳性率为 15%）为测试队列。AI 系统对测试队列的 AUC 为 0.74（95%置信区间 [CI] 0.58-0.86），而使用指南标准的 AUC 为 0.52（95% CI 0.50-0.55）（P=0.0028）。与指南相比，该 AI 模型可减少 21%的过度手术。