OBJECTIVE To develop a machine learning-based ultrasound (US) radiomics model for predicting tumour deposits (TDs) preoperatively. METHODS From December 2015 to December 2017, 127 patients with rectal cancer were prospectively enrolled and divided into training and validation sets. Endorectal ultrasound (ERUS) and shear-wave elastography (SWE) examinations were conducted for each patient. A total of 4176 US radiomics features were extracted for each patient. After the reduction and selection of US radiomics features , a predictive model using an artificial neural network (ANN) was constructed in the training set. Furthermore, two models (one incorporating clinical information and one based on MRI radiomics) were developed. These models were validated by assessing their diagnostic performance and comparing the areas under the curve (AUCs) in the validation set. RESULTS The training and validation sets included 29 (33.3%) and 11 (27.5%) patients with TDs, respectively. A US radiomics ANN model was constructed. The model for predicting TDs showed an accuracy of 75.0% in the validation cohort. The sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV) and AUC were 72.7%, 75.9%, 53.3%, 88.0% and 0.743, respectively. For the model incorporating clinical information, the AUC improved to 0.795. Although the AUC of the US radiomics model was improved compared with that of the MRI radiomics model (0.916 vs. 0.872) in the 90 patients with both ultrasound and MRI data (which included both the training and validation sets), the difference was nonsignificant (p = 0.384). CONCLUSIONS US radiomics may be a potential model to accurately predict TDs before therapy. KEY POINTS • We prospectively developed an artificial neural network model for predicting tumour deposits based on US radiomics that had an accuracy of 75.0%. • The area under the curve of the US radiomics model was improved than that of the MRI radiomics model (0.916 vs. 0.872), but the difference was not significant (p = 0.384). • The US radiomics-based model may potentially predict TDs accurately before therapy, but this model needs further validation with larger samples.

中文翻译：

---

基于人工神经网络的美国放射组学模型对直肠癌肿瘤沉积的术前预测。

目的 开发一种基于机器学习的超声 (US) 放射组学模型，用于术前预测肿瘤沉积 (TD)。方法 从 2015 年 12 月到 2017 年 12 月，前瞻性纳入 127 名直肠癌患者，并将其分为训练集和验证集。对每位患者进行了直肠内超声 (ERUS) 和剪切波弹性成像 (SWE) 检查。为每位患者提取了总共 4176 个美国放射组学特征。在对美国放射组学特征进行缩减和选择后，在训练集中构建了使用人工神经网络 (ANN) 的预测模型。此外，还开发了两种模型（一种包含临床信息，另一种基于 MRI 放射组学）。这些模型通过评估其诊断性能并比较验证集中的曲线下面积 (AUC) 来验证。结果 训练集和验证集分别包括 29 名 (33.3%) 和 11 名 (27.5%) 的 TD 患者。构建了美国放射组学 ANN 模型。预测 TD 的模型在验证队列中的准确率为 75.0%。敏感性、特异性、阳性预测值（PPV）、阴性预测值（NPV）和AUC分别为72.7%、75.9%、53.3%、88.0%和0.743。对于包含临床信息的模型，AUC 提高到 0.795。尽管在具有超声和 MRI 数据（包括训练集和验证集）的 90 名患者中，与 MRI 放射组学模型相比，美国放射组学模型的 AUC 有所提高（0.916 对 0.872），但 差异不显着 (p = 0.384)。结论 美国放射组学可能是一种在治疗前准确预测 TD 的潜在模型。要点 • 我们前瞻性地开发了一种基于美国放射组学的预测肿瘤沉积的人工神经网络模型，其准确率为 75.0%。• 美国放射组学模型的曲线下面积比MRI 放射组学模型有所改善（0.916 对0.872），但差异不显着（p = 0.384）。• 美国基于放射组学的模型可能会在治疗前准确预测 TD，但该模型需要用更大的样本进一步验证。要点 • 我们前瞻性地开发了一种基于美国放射组学的预测肿瘤沉积的人工神经网络模型，其准确率为 75.0%。• 美国放射组学模型的曲线下面积比MRI 放射组学模型有所改善（0.916 对0.872），但差异不显着（p = 0.384）。• 美国基于放射组学的模型可能会在治疗前准确预测 TD，但该模型需要用更大的样本进一步验证。要点 • 我们前瞻性地开发了一种基于美国放射组学的预测肿瘤沉积的人工神经网络模型，其准确率为 75.0%。• 美国放射组学模型的曲线下面积比MRI 放射组学模型有所改善（0.916 对0.872），但差异不显着（p = 0.384）。• 美国基于放射组学的模型可能会在治疗前准确预测 TD，但该模型需要用更大的样本进一步验证。