多参数MRI影像组学术前鉴别睾丸精原细胞瘤与非精原细胞瘤

RATIONALE AND OBJECTIVES: This study aims to develop an optimal model for distinguishing seminoma from non-seminoma testicular tumors using machine learning classifiers based on multiparametric MRI radiomics.

MATERIALS AND METHODS: This multi-institutional study enrolled a total of 188 patients, including 83 with seminoma and 105 with non-seminoma. The cohort from Institution 1 (n = 137) served as the training and validation set, whereas the independent cohort from Institution 2 (n = 51) was designated as the test set. Manual segmentation of tumor regions of interest (ROIs) was performed by experienced researchers on DWI, ADC, T2WI, and CE-T1WI sequences. A comprehensive radiomics workflow was implemented, encompassing data standardization, dimensionality reduction, feature selection, and classification using six distinct machine learning classifiers. Predictive models were developed by integrating radiomics features extracted from individual sequences and multiple sequence combinations with machine learning algorithms. In parallel, clinical models were established through univariate and multivariate logistic regression to identify significant predictive factors. A combined model incorporating both radiomics signatures and clinical characteristics was subsequently developed. The discriminatory performance of all models was evaluated by comparing area under the curve (AUC) values using DeLong's test.

RESULTS: Model 19 (DWI+T2WI+ADC+CE-T1W+Clinical/SVM) achieved the highest AUC values, reaching 0.934, 0.922, and 0.846 on the training, validation, and external test sets respectively. In terms of clinical-radiological characteristics, alpha-fetoprotein (AFP) and cystic necrosis are significant predictors.

CONCLUSION: A combined model utilising DWI, ADC, T2WI, and contrast-enhanced T1WI imaging, alongside AFP and cystic necrosis, demonstrates high precision, adaptability, and robustness in distinguishing seminoma from non-seminoma testicular tumors.