Huang Qiao-Yi, Zheng Hui-da, Xiong Bin, Huang Qi-Ming, Ye Kai, Lin Shu, Xu Jian-Hua
Department of Gynaecology and Obstetrics, The Second Affiliated Hospital, Fujian Medical University, Quanzhou, Fujian Province, China.
Department of Gastrointestinal Surgery, The Second Affiliated Hospital of Fujian Medical University, Quanzhou, Fujian Province, China.
Heliyon. 2025 Jan 9;11(2):e41852. doi: 10.1016/j.heliyon.2025.e41852. eCollection 2025 Jan 30.
Colorectal cancer (CRC) is the second most prevalent cause of oncological mortality, and its diagnostic and therapeutic decision-making processes is complex. Alteration in molecular characteristic expression is closely related to tumor invasiveness and can serve a novel biomarker for predicting cancer prognosis. In this study, we aimed to construct radiomic models through machine learning to predict the progression of CRC. We collected the clinical, pathological, and magnetic resonance imaging (MRI) data of 136 CRC patients who underwent direct surgical resection. Immunohistochemistry analysis was performed to detect the expression levels of p53, synaptophysin (Syn), human epidermal growth factor receptor 2 (HER2), perineural invasion (PNI), and vascular invasion (VI) expression levels in CRC tissues. After the manual lesion segmentation, 1781 radiomics features were extracted from the transverse T2-weighted image of MRI (T2W-MRI). We employed Spearman's rank correlation coefficient, greedy recursive deletion strategy, minimum redundancy, maximum relevance, least absolute shrinkage, and selection operator regression were utilized to screen for radiological features. Radiomics and clinical models were constructed using the K-nearest neighbor (KNN). The diagnostic efficiencies of the prediction models were evaluated using receiver operating characteristic curves and quantified employing the area under the curve (AUC). Our research results indicate that compared with the single radioactive model, the clinical radiomics model in the validation cohort showed better diagnostic performance, as indicated by the AUC values (p53 = 0.758, Syn = 0.739, HER2 = 0.786, PNI = 0.835, VI = 0.797). Furthermore, the calibration curve and decision curve analyses showed the clinical benefits. In summary, we developed and validated a clinical radiomics model to preoperative prediction of the biological characteristic expression levels of CRC. The findings of this research may offer a promising noninvasive method for evaluating CRC risk stratification and may lay the groundwork for treatment of this disease.
结直肠癌（CRC）是肿瘤学死亡率的第二大常见原因，其诊断和治疗决策过程复杂。分子特征表达的改变与肿瘤侵袭性密切相关，可作为预测癌症预后的新型生物标志物。在本研究中，我们旨在通过机器学习构建放射组学模型来预测CRC的进展。我们收集了136例行直接手术切除的CRC患者的临床、病理和磁共振成像（MRI）数据。进行免疫组织化学分析以检测CRC组织中p53、突触素（Syn）、人表皮生长因子受体2（HER2）、神经周围浸润（PNI）和血管浸润（VI）的表达水平。在手动病变分割后，从MRI的横向T2加权图像（T2W-MRI）中提取了1781个放射组学特征。我们采用Spearman等级相关系数、贪婪递归删除策略、最小冗余、最大相关性、最小绝对收缩和选择算子回归来筛选放射学特征。使用K近邻（KNN）构建放射组学和临床模型。使用受试者工作特征曲线评估预测模型的诊断效率，并通过曲线下面积（AUC）进行量化。我们的研究结果表明，与单一放射性模型相比，验证队列中的临床放射组学模型显示出更好的诊断性能，AUC值表明了这一点（p53 = 0.758，Syn = 0.739，HER2 = 0.786，PNI = 0.835，VI = 0.797）。此外，校准曲线和决策曲线分析显示了临床益处。总之，我们开发并验证了一种临床放射组学模型用于术前预测CRC的生物学特征表达水平。本研究结果可能为评估CRC风险分层提供一种有前景的非侵入性方法，并可能为该疾病的治疗奠定基础。
