Department of Radiology, The Second Affiliated Hospital of Soochow University, Suzhou, China.
Department of MRI Room, Yancheng First Hospital Affiliated Hospital of NanJing University Medical School, Yancheng, China.
BMC Med Imaging. 2024 Jul 25;24(1):185. doi: 10.1186/s12880-024-01372-8.
Exploring the value of adding correlation analysis (radiomic features (RFs) of pelvic metastatic lymph nodes and primary lesions) to screen RFs of primary lesions in the feature selection process of establishing prediction model.
A total of 394 prostate cancer (PCa) patients (263 in the training group, 74 in the internal validation group and 57 in the external validation group) from two tertiary hospitals were included in the study. The cases with pelvic lymph node metastasis (PLNM) positive in the training group were diagnosed by biopsy or MRI with a short-axis diameter ≥ 1.5 cm, PLNM-negative cases in the training group and all cases in validation group were underwent both radical prostatectomy (RP) and extended pelvic lymph node dissection (ePLND). The RFs of PLNM-negative lesion and PLNM-positive tissues including primary lesions and their metastatic lymph nodes (MLNs) in the training group were extracted from T2WI and apparent diffusion coefficient (ADC) map to build the following two models by fivefold cross-validation: the lesion model, established according to the primary lesion RFs selected by t tests and absolute shrinkage and selection operator (LASSO); the lesion-correlation model, established according to the primary lesion RFs selected by Pearson correlation analysis (RFs of primary lesions and their MLNs, correlation coefficient > 0.9), t test and LASSO. Finally, we compared the performance of these two models in predicting PLNM.
The AUC and the DeLong test of AUC in the lesion model and lesion-correlation model were as follows: training groups (0.8053, 0.8466, p = 0.0002), internal validation group (0.7321, 0.8268, p = 0.0429), and external validation group (0.6445, 0.7874, p = 0.0431), respectively.
The lesion-correlation model established by features of primary tumors correlated with MLNs has more advantages than the lesion model in predicting PLNM.
探讨在建立预测模型的特征选择过程中，将相关性分析（盆腔转移性淋巴结和原发灶的放射组学特征（RFs））加入到原发灶 RFs 中筛选的价值。
本研究共纳入来自两家三级医院的 394 例前列腺癌（PCa）患者（训练组 263 例，内部验证组 74 例，外部验证组 57 例）。训练组中经活检或 MRI 诊断为盆腔淋巴结转移（PLNM）阳性的病例，其短轴直径≥1.5cm；训练组中 PLNM 阴性病例和验证组中所有病例均接受根治性前列腺切除术（RP）和扩展盆腔淋巴结清扫术（ePLND）。从 T2WI 和表观扩散系数（ADC）图中提取训练组中 PLNM 阴性病灶和 PLNM 阳性组织（包括原发灶及其转移性淋巴结（MLNs））的 RFs，通过五重交叉验证构建以下两种模型：根据 t 检验和绝对收缩选择算子（LASSO）选择的原发灶 RFs 构建的病灶模型；根据 Pearson 相关性分析（原发灶 RFs 及其 MLNs，相关系数>0.9）、t 检验和 LASSO 选择的原发灶 RFs 构建的病灶相关模型。最后，比较这两种模型预测 PLNM 的性能。
病灶模型和病灶相关模型的 AUC 和 DeLong 检验 AUC 在训练组（0.8053、0.8466，p=0.0002）、内部验证组（0.7321、0.8268，p=0.0429）和外部验证组（0.6445、0.7874，p=0.0431）中的表现如下。
与 MLNs 相关的原发肿瘤特征建立的病灶相关模型在预测 PLNM 方面优于病灶模型。
