目的:研究基于增强CT的放射组学模型在直肠癌术前多分类T分期诊断中的可行性和有效性。方法:本研究纳入了500例符合条件的直肠癌患者(T1期为48例,T2期为127例,T3期为259例,T4期为66例),并随机分为训练组(n = 400例)和验证组(n = 100例)。直肠癌病变的感兴趣区由资深放射科医生进行像素化标注。我们使用pyradiomics从人工分割的感兴趣区域中提取特征,然后采取10种不同机器学习算法构建了放射组学多分类模型,并使用这10种算法构建临床基线数据多分类模型。使用ROC曲线、micro-average AUC、macro-average AUC和准确率评估多分类模型的性能。结果:10种放射组学多分类模型中预测效果最好的Model SVM (micro-average AUC = 0.845,macro-average AUC = 0.777,准确率 = 0.600)和10种临床基线数据多分类模型中预测效果最好的Model SVM (micro-average AUC = 0.841,macro-average AUC = 0.760,准确率 = 0.610)相比,在进行直肠癌T1、T2、T3和T4分期诊断方面表现更好。结论:综上所述,与临床基线数据多分类模型相比,基于CT的放射组学多分类模型在进行直肠癌分期诊断方面具有更优异的预测性能。Purpose: To investigate the feasibility and efficacy of a contrast-enhanced CT-based radiomics model in the preoperative multi-class T-staging diagnosis of rectal cancer. Methods: This study enrolled 500 eligible rectal cancer patients (48 T1, 127 T2, 259 T3, and 66 T4 cases) and randomly divided them into a training cohort (n = 400) and a validation cohort (n = 100). Regions of interest (ROIs) for rectal cancer lesions were manually segmented by senior radiologists. Using Pyradiomics, we extracted features from the annotated ROIs and constructed multi-class radiomics models with 10 machine learning algorithms. Additionally, clinical baseline models were developed using the same algorithms. Model performance was evaluated using ROC curves, micro-average AUC, macro-average AUC, and accuracy. Results: Among t
