摘要
Purpose:
Lymphovascular space invasion (LVSI) is an independent risk factor for lymph node metastasis in cervical cancer and is associated with poor prognosis, highlighting the need for early detection to optimize treatment. Therefore, this study aims to investigate the potential of multi-sequence MRI habitat analysis in predicting LVSI in cervical cancer patients.
Methods：
This retrospective study included 52 patients with pathologically confirmed cervical squamous cell carcinoma(16 LVSI-positive, 36 LVSI-negative). All patients underwent pelvic MRI on a 3.0T scanner with a 48-channel phased-array coil. The protocol included a T1-weighted high-resolution contrast-enhanced (T1-HRCE) sequence(TR/TE= 4.83/2.37 ms, FOV=240×240 mm2, matrix=352×352) and an intravoxel incoherent motion(IVIM) sequence with 11 b-values: 01, 201, 401, 501, 801, 1001, 2001, 5002, 8003, 15006, 20006 s/mm2(TR/TE= 2800/83.2 ms, FOV=240×240 mm2, matrix=128×128). Tumor volumes of interest (VOIs) were manually delineated on T1-HRCE images using 3D Slicer(v5.6.1; https://www.slicer.org). Based on voxel-wise analysis of T1-HRCE and IVIM-derived parameters (pure diffusion coefficient [D] and perfusion fraction [F]), tumors were partitioned into three distinct subregions (Habitat1, Habitat2, and Habitat3) via K-means clustering. Subsequently, Radiomics features were extracted from each subregion, followed by sequential selection of significant features: group-differentiating features were first screened using the Mann-Whitney U test (p < 0.1); highly correlated features were then removed via Spearman’s correlation analysis (|r| > 0.9); and finally, the optimal penalty coefficient for LASSO regression was determined through five-fold cross-validation to complete feature selection. A forward stepwise logistic regression model was then constructed of each subregion to predict LVSI status. Model performance was assessed using the area under the receiver operating characteristic curve (AUC), calibration curve, and additional metrics including accuracy, sensitivity, specificity, F1-score, Brier score and Hosmer-Lemeshow test statistics.
Results:
From an initial set of 321 radiomics features, 5, 5, and 4 features were retained for Habitat 1, 2 and 3, respectively. The Habitat2 model showed the highest predictive performance, with an AUC of 0.870(95% CI:0.770-0.970), accuracy of 0.827, sensitivity of 0.812, specificity of 0.833, F1-score of 0.743. Additionally, the calibration curve of the Habitat2 model demonstrated satisfactory consistency between predicted and observed outcomes, supported by a Brier score of 0.134 and a Hosmer-Lemeshow test statistic of 0.571(χ2=6.684, df=8).
Conclusions:
Our subregion-specific radiomics models effectively quantified tumor heterogeneity and demonstrated strong discriminatory performance in predicting LVSI status in cervical cancer. Notably, Habitat2, characterized by diminished vascularity and cellularity (low F and intermediate-to-high D), showed the strongest association with LVSI status, suggesting a potential link to higher necrosis activity within this tumor subregion.
