Related Experiment Video
Updated: Jan 9, 2026
Author Spotlight: Advancing Early Detection and Treatment of Gastrointestinal Tumors
Published on: February 16, 2024
Deep Learning and Attention Mechanism-based Prediction of Vaginal Invasion in Early-Stage Cervical Cancer
Qing Xu1, Chao He2, Xinyang Zhu3
1Department of Medical Techniques, Shaanxi University of Chinese Medicine, Shaanxi, 712046, China.
Introduction:
This study introduces a novel fusion of 3D ResNet classification and Grad-CAM visualization to predict vaginal invasion in early-stage cervical cancer using T2WI-MRI, enhancing diagnostic accuracy while enabling anatomical localization of invasive lesions.
Methods:
This retrospective study analyzed sagittal T2WI from 160 patients with pathologically confirmed stage IB-IIA cervical cancer to predict vaginal invasion. Following an 8:2 training-test split, radiomic features were extracted from manually delineated intratumoral regions and four concentrically expanded peritumoral regions (1-4mm). Features selection by Pearson correlation and LASSO regression. Random forest models incorporating intratumoral and peritumoral (0-4mm) features were constructed, with ROC analysis identifying the optimal model. Subsequently, a 3D-ResNet architecture, enhanced with anisotropic convolutional layers and sophisticated data augmentation, was developed and optimized using the optimal ROI configuration. Model interpretability was facilitated using Grad-CAM, with performance assessed by AUC, sensitivity, specificity, accuracy, and precision.
Results:
The AIC-enhanced 3D ResNet-18 model, integrating intratumoral and 3mm peritumoral regions, showed superior test performance (AUC: 0.784, Sensitivity: 0.650, Specificity: 0.765, Accuracy: 0.611, Precision: 0.686) versus the baseline (AUC: 0.742), representing a 6% AUC improvement. Grad-CAM heatmaps identified diagnostically relevant regions within the tumor microenvironment, enhancing biological plausibility and model interpretability.
Discussion:
This attention-integrated 3D ResNet-18 framework (AUC=0.784) facilitates non-invasive vaginal invasion detection for fertility-sparing decisions, validated through Grad-CAM tumor localization; however, derivation from a single-center cohort warrants external validation and prospective studies before clinical translation.
Conclusion:
This preliminary study demonstrates promising deep learning performance (3D ResNet-18+Grad-CAM+AIC) for vaginal invasion assessment, despite moderate n; however, a single-center retrospective design limits generalizability.
