基于双参数MRI影像组学诊断前列腺癌及评估危险分级的研究

Research on the Diagnosis of Prostate Cancer and Evaluation of Risk Grading Based on Biparametric MRI Radiomics

摘要: 目的：前列腺癌是男性高发的恶性肿瘤，早期准确诊断、危险分层是制定治疗方案、提高患者预后的关键。双参数MRI (bpMRI)把T2加权成像(T2WI)和弥散加权成像(ADC)结合起来，可以无创地反映前列腺病灶的解剖形态和水分子扩散特性，已经广泛应用于前列腺癌的筛查和定位。影像组学属于新兴技术，可以从医学影像当中找出人类视觉无法察觉的量化特征，给疾病诊断以及预后评价赋予客观的参照。方法：回顾性收集前列腺穿刺或者全切术后病理确诊为前列腺癌患者213例，获得患者的Gleason评分、临床分期、血清PSA值、影像资料。使用ITK-SNAP软件手动勾画病灶感兴趣区(ROI)，用Pyradiomics工具提取形态、灰度、纹理等维度的影像组学特征，用Spearman相关性检验、LASSO、PCA等方法降维筛选有效特征。本研究采用十折交叉验证策略，对训练集样本进行随机划分为10个子集，并基于筛选出的关键特征对分类器进行迭代优化，以完成影像组学特征的筛选工作。随后，分别基于T2WI、ADC以及T2WI与ADC融合序列构建六种不同的机器学习模型，并对所构建的模型进行全面评估，系统分析各模型的准确率、特异度及灵敏度等性能指标。结果：根据T2WI序列建立的六个模型的验证集AUC分别为SVM (0.793)、KNN (0.707)、RF (0.795)、ET (0.773)、XGBoost (0.758)、LightGBM (0.822)。基于ADC序列建立的六个模型的验证集AUC分别为SVM (0.857)、KNN (0.748)、RF (0.733)、ET (0.818)、XGBoost (0.789)、LightGBM (0.825)。基于T2WI + ADC组合模型的验证集AUC分别为SVM (0.855)、KNN (0.771)、RF (0.781)、ET (0.850)、XGBoost (0.893)、LightGBM (0.859)。其中T2WI加ADC组合序列的XGBoost模型有最好的诊断效能。结论：基于磁共振多参数影像组学技术与机器学习算法的融合分析，能够对病灶的影像学特征进行量化评估，从而实现对中低危与高危前列腺癌的有效鉴别诊断，实现前列腺癌的精准诊断和危险分级，临床与影像组学联合模型可以进一步提高预测的准确性，为减少不必要的穿刺、制定个体化的治疗方案提供可靠的无创评估工具。

Abstract: Objective: Prostate cancer is a highly prevalent malignant tumor in men. Early and accurate diagnosis and risk stratification are crucial for formulating treatment plans and improving patient prognosis. Biparametric MRI (bpMRI), which combines T2-weighted imaging (T2WI) and diffusion-weighted imaging (ADC), can non-invasively reflect the anatomical morphology and water molecule diffusion characteristics of prostate lesions and has been widely used in the screening and localization of prostate cancer. Radiomics is an emerging technology that can identify quantitative features undetectable by human vision from medical images, providing objective references for disease diagnosis and prognosis evaluation. Methods: A retrospective collection of 213 patients with prostate cancer confirmed by prostate biopsy or radical prostatectomy was conducted. The patients’ Gleason scores, clinical stages, serum PSA values, and imaging data were obtained. The region of interest (ROI) of the lesion was manually delineated using ITK-SNAP software, and radiomics features in terms of morphology, gray level, and texture were extracted using the Pyradiomics tool. Dimensionality reduction and feature selection were performed using Spearman correlation test, LASSO, and PCA methods. This study adopted a ten-fold cross-validation strategy, randomly dividing the training set samples into 10 subsets, and iteratively optimizing the classifier based on the selected key features to complete the radiomics feature selection. Subsequently, six different machine learning models were constructed based on T2WI, ADC, and the fusion sequence of T2WI and ADC, and the constructed models were comprehensively evaluated to systematically analyze the performance indicators such as accuracy, specificity, and sensitivity of each model. Results: The AUC values of the six models based on the T2WI sequence in the validation set were SVM (0.793), KNN (0.707), RF (0.795), ET (0.773), XGBoost (0.758), and LightGBM (0.822). The AUC values of the six models based on the ADC sequence in the validation set were SVM (0.857), KNN (0.748), RF (0.733), ET (0.818), XGBoost (0.789), and LightGBM (0.825). The AUC values of the six models based on the T2WI + ADC combined sequence in the validation set were SVM (0.855), KNN (0.771), RF (0.781), ET (0.850), XGBoost (0.893), and LightGBM (0.859). Among them, the XGBoost model based on the T2WI + ADC combined sequence had the best diagnostic performance. Conclusion: The integration of multi-parameter MRI radiomics technology and machine learning algorithms can quantitatively evaluate the imaging features of lesions, thereby achieving effective differential diagnosis of intermediate-low-risk and high-risk prostate cancer, and realizing precise diagnosis and risk stratification of prostate cancer. The combination of clinical and radiomics models can further improve the accuracy of prediction, providing a reliable non-invasive assessment tool for reducing unnecessary biopsies and formulating individualized treatment plans.