Comparative evaluation of conventional radiomics and VGG-SAM fusion strategies for MRI-based preoperative prediction of perineural invasion in cervical cancer.
Authors
Affiliations (4)
Affiliations (4)
- Department of Gynecologic Oncology, National Cancer Center/National Clinical Research Center for Cancer/Cancer Hospital, Chinese Academy of Medical Sciences and Peking Union Medical College, Beijing, China.
- Department of Orthopedics, The Fourth Medical Center of Chinese PLA General Hospital, Beijing, China.
- Academy of Artificial Intelligence, Beijing Institute of Petrochemical Technology, Beijing, China. [email protected].
- Department of Gynecologic Oncology, National Cancer Center/National Clinical Research Center for Cancer/Cancer Hospital, Chinese Academy of Medical Sciences and Peking Union Medical College, Beijing, China. [email protected].
Abstract
Perineural invasion (PNI) is an adverse feature in cervical cancer and may influence nerve-sparing surgery. We compared conventional radiomics and VGG-SAM deep-learning strategies for MRI-based preoperative prediction of PNI. A retrospective cohort of 103 patients with cervical cancer, including 82 PNI-negative and 21 PNI-positive patients, was analyzed. PyRadiomics features were extracted from raw DICOM MRI and paired tumor masks. Four deep variants were evaluated: VGG only, SAM only, VGG + SAM naive fusion, and VGG + SAM learnable fusion. Patient-level stratified five-fold cross-validation was used. Primary metrics were sensitivity, specificity, balanced accuracy, F1-score, ROC-AUC, and PR-AUC. Uncertainty for the naive-versus-learnable comparison was assessed with 10,000 paired bootstrap resamples and exact McNemar tests. The strongest radiomics baseline, PyRadiomics + XGBoost, achieved balanced accuracy 0.6057, F1-score 0.3684, ROC-AUC 0.6731, and PR-AUC 0.4460. Naive VGG + SAM fusion achieved the best sensitivity (0.5714), balanced accuracy (0.7491), F1-score (0.6154), ROC-AUC (0.7854), and PR-AUC (0.6231). Learnable fusion achieved the highest specificity (0.9634) and precision (0.7500) while matching the highest accuracy (0.8544). Paired bootstrap comparisons showed wide confidence intervals across clinically relevant metrics. Hybrid VGG-SAM modeling outperformed conventional radiomics alone, but greater fusion complexity did not yield uniform benefit. Naive fusion favored sensitivity-oriented discrimination, whereas learnable fusion mainly shifted the operating point toward higher specificity. Larger multi-institutional cohorts and clinical comparators are needed.