Amide proton transfer-weighted habitat radiomics: a superior approach for preoperative prediction of lymphovascular space invasion in cervical cancer

TL;DR

This study shows that using whole-tumor APTw habitat radiomics improves the preoperative prediction of lymphovascular space invasion in cervical cancer compared to traditional methods.

Contribution

The study introduces whole-tumor APTw habitat radiomics as a novel method for predicting lymphovascular space invasion in cervical cancer.

Findings

The APTw_h3 model achieved an AUC of 0.796 for predicting LVSI positivity.

Combining APTw_h3 with clinical and radiological features improved AUC to 0.903.

Adding APTw_h3 increased sensitivity and specificity for determining LVSI positivity.

Abstract

Non-invasive preoperative prediction of lymphovascular space invasion (LVSI) in cervical cancer (CC) is clinically important for guiding surgical planning and adjuvant therapy, while avoiding the risks associated with invasive procedures. However, current studies using amide proton transfer-weighted (APTw) MRI for LVSI prediction typically analyze only the mean values from a limited number of intratumoral regions of interest (ROIs), which fails to fully capture tumor heterogeneity. This study investigates the added value of whole-tumor APTw habitat radiomics in predicting LVSI and its advantages over conventional analysis methods.

This prospective study included consecutive adult patients with suspected CC who underwent APTw MRI between December 2022 and December 2024; a portion of the cohort has been reported previously. APTw values were extracted using two methods: (1) the conventional approach, calculating the mean signal from three ROIs on a representative slice; and (2) habitat radiomics, involving whole-tumor segmentation, k-means clustering to identify functional subregions, and radiomic feature extraction. Pathological assessment of LVSI from hysterectomy specimens served as the reference standard. Multivariable logistic regression identified variables associated with LVSI and developed diagnostic models. Model robustness was evaluated by 5-fold cross-validation, with AUC and DeLong’s test used for performance assessment.

Among 124 patients (74 LVSI−, 50 LVSI+), the APTw_h3 model achieved a higher AUC (0.796 [95% CI: 0.709–0.882]) for predicting LVSI positivity than the clinical-radiological model (AUC = 0.733, 95% CI: 0.638–0.817). The combined model integrating clinical, radiological, and APTw_h3 features achieved the highest AUC (0.903, 95% CI: 0.841–0.952), which was significantly higher than those of both the clinical-radiological and APTw_h3 models (both P < 0.001). Moreover, the addition of APTw_h3 to the clinical-radiological model improved sensitivity (88% vs. 82%) and specificity (83.8% vs. 64.9%) for determining LVSI positivity.

Whole-tumor APTw habitat radiomics demonstrates superior performance over conventional mean-value APTw analysis for preoperative prediction of LVSI in CC. Notably, integrating habitat radiomic features with clinical and radiological parameters further improves predictive accuracy, demonstrating potential for enhanced individualized patient management.

Figures