Abstract：
Risk evaluation of lymph node metastasis (LNM) for endoscopically resected submucosal invasive (T1) colorectal cancers (CRC) is critical for determining therapeutic strategies, but interobserver variability for histologic evaluation remains a major problem. To address this issue, we developed a machine-learning model for predicting LNM of T1 CRC without histologic assessment. A total of 783 consecutive T1 CRC cases were randomly split into 548 training and 235 validation cases. First, we trained convolutional neural networks (CNN) to extract cancer tile images from whole-slide images, then re-labeled these cancer tiles with LNM status for re-training. Statistical parameters of the tile images based on the probability of primary endpoints were assembled to predict LNM in cases with a random forest algorithm, and defined its predictive value as random forest score. We evaluated the performance of case-based prediction models for both training and validation datasets with area under the receiver operating characteristic curves (AUC). The accuracy for classifying cancer tiles was 0.980. Among cancer tiles, the accuracy for classifying tiles that were LNM-positive or LNM-negative was 0.740. The AUCs of the prediction models in the training and validation sets were 0.971 and 0.760, respectively. CNN judged the LNM probability by considering histologic tumor grade.
摘要：
内镜切除的粘膜下浸润性（T1）结直肠癌（CRC）淋巴结转移（LNM）的风险评估对于确定治疗策略至关重要。但是组织学评估的观察者间差异仍然是一个主要问题。为了解决这个问题，我们开发了一种机器学习模型,用于在不进行组织学评估的情况下预测T1CRC的LNM.总共783例连续的T1CRC病例被随机分为548例训练和235例验证病例。首先,我们训练卷积神经网络（CNN）从整张幻灯片图像中提取癌症图像，然后用LNM状态重新标记这些癌症瓷砖以进行重新训练。在随机森林算法的情况下，基于主要终点概率的图块图像的统计参数被组合以预测LNM。并将其预测值定义为随机森林得分。我们评估了基于病例的预测模型在受试者工作特征曲线（AUC）下的训练和验证数据集的性能。癌症瓷砖分类的准确性为0.980。在癌症瓷砖中，对LNM阳性或LNM阴性的瓷砖进行分类的准确性为0.740.训练集和验证集中预测模型的AUC分别为0.971和0.760。CNN通过考虑组织学肿瘤分级来判断LNM概率。
