基于深度学习的放射影像组学特征预测局部晚期直肠癌新辅助放化疗反应研究

Objective To explore the effectiveness of deep learning-based methods in extracting radiomics features from pre-treatment MRI to predict the response to neoadjuvant chemoradiotherapy in locally advanced rectal cancer. Methods From 2016 to 2017, 43 patients with locally advanced rectal cancer who received neoadjuvant concurrent chemoradiotherapy were included. All patients underwent total mesorectal excision 6 - 12 weeks after treatment. Diffusion-weighted imaging (DWI) sequence MRI was obtained before concurrent chemoradiotherapy. According to postoperative pathology, imaging examination or colonoscopy, the patients were divided into a treatment response group (22 cases) and a treatment non-response group (21 cases). Traditional computer-aided diagnosis methods and a pre-trained convolutional neural network were used respectively to extract manual and deep learning-based radiomics (DLR) features from the apparent diffusion coefficient maps of the DWI sequence. The least absolute shrinkage and selection operator Logistic regression model was established using the extracted features to predict the treatment response. The receiver operating characteristic curve was used to evaluate the model performance through 20 repetitions of stratified 4-fold cross-validation. Results The average area under the curve of the model constructed using DLR was 0.73 (standard error 0.58 - 0.80). Conclusion The radiomics features extracted from pre-treatment MRI using deep learning methods have high accuracy in predicting the neoadjuvant treatment response in patients with locally advanced rectal cancer.

目的探讨基于深度学习的方法,从疗前MRI中提取放射影像组学特征预测局部晚期直肠癌新辅助放化疗反应的有效性。方法2016-2017年纳入43例局部晚期直肠癌新辅助同步放化疗患者。均在疗后6~12周接受全系膜直肠切除术。弥散加权成像(DWI)序列MRI在同步放化疗前获得。根据术后病理、影像学检查或肠镜检查评估新辅助治疗后反应,将患者分为治疗反应组(22例)和治疗无反应组(21例)。分别采用传统的计算机辅助诊断方法和预先训练的卷积神经网络,从DWI序列的表观扩散系数图中提取手工和基于深度学习的影像组学(DLR)特征。利用提取的特征建立最小绝对收缩和选择算子Logistic回归模型,预测治疗反应。使用受试者工作特性曲线,通过重复20次分层4倍交叉验证评估模型性能。结果使用基于DLR构建模型的平均曲线下面积为0.73(标准误为0.58~0.80)。结论从疗前MRI中基于深度学习方法提取的影像组学特征在预测局部晚期直肠癌患者新辅助治疗反应方面准确度高。