Deep Learning and Attention Mechanism-based Prediction of Vaginal Invasion in Early-Stage Cervical Cancer

Area of Science:

• Oncology
• Radiology
• Artificial Intelligence

Background:

• Early-stage cervical cancer diagnosis requires accurate assessment of vaginal invasion for appropriate treatment.
• Current diagnostic methods may have limitations in precisely identifying the extent of tumor invasion.

Purpose of the Study:

• To develop and validate a novel deep learning model for predicting vaginal invasion in early-stage cervical cancer using T2WI-MRI.
• To enhance diagnostic accuracy and provide anatomical localization of invasive lesions.

Main Methods:

• A retrospective analysis of 160 patients with stage IB-IIA cervical cancer using sagittal T2WI-MRI.
• Extraction of radiomic features from intratumoral and peritumoral regions, followed by feature selection.
• Development of an attention-integrated 3D ResNet-18 model, optimized with anisotropic convolutional layers and data augmentation.
• Model interpretability and localization of invasive regions using Grad-CAM visualization.

Main Results:

• The AIC-enhanced 3D ResNet-18 model achieved an AUC of 0.784, outperforming the baseline model (AUC=0.742) with a 6% AUC improvement.
• Grad-CAM heatmaps successfully identified diagnostically relevant regions within the tumor microenvironment, enhancing interpretability.
• The model demonstrated good performance in sensitivity (0.650), specificity (0.765), accuracy (0.611), and precision (0.686).

Conclusions:

• The developed 3D ResNet-18 framework with Grad-CAM visualization shows promise for non-invasive detection of vaginal invasion, supporting fertility-sparing treatment decisions.
• The AI model's ability to localize tumors enhances biological plausibility and clinical relevance.
• External validation and prospective studies are recommended due to the single-center retrospective design.