An Explainable Deep Learning Framework for Predicting Postoperative Radiotherapy-Induced Vaginal Stenosis in Surgically Treated Cervical Cancer Patients.

Abstract

Aim: Surgery (e.g., radical hysterectomy) combined with radiotherapy is the mainstay of treatment strategy for locally advanced cervical cancer. However, the beneficial effects of adjuvant radiotherapy are frequently offset by late-onset toxicities, such as vaginal stenosis (VS), which significantly impact patients' quality of life. Although imaging techniques like computed tomography (CT) and magnetic resonance imaging (MRI) are key for both surgical planning and radiotherapy targeting, their ability to predict VS risk before treatment remains limited. This challenge underscores the need for accurate and interpretable predictive models specifically adapted to surgical oncology contexts. This study aims to develop and validate an explainable deep learning framework, integrating Squeeze-and-Excitation (SE) networks and Gradient-weighted Class Activation Mapping (Grad-CAM) visualization, for predicting radiotherapy-induced VS to enable early, personalized intervention strategies.

Methods: Pre-treatment (i.e., post-surgical, pre-radiotherapy) CT images of cervical cancer patients diagnosed between January 2017 and March 2022 were retrospectively collected. These patients underwent radical hysterectomy (or equivalent surgical resection) followed by radiotherapy. Each patient was categorized as either positive or negative for subsequent VS development. Following normalization and augmentation, we employed a Squeeze-and-Excitation enhanced Inception network (SE-Inception) to distinguish between high- and low-risk cases. Model performance was compared to a conventional Random Forest and a deep learning baseline (ResNet50). Additionally, Grad-CAM visualization was integrated to highlight discriminative image regions for enhanced interpretability and clinical validation.

Results: Among the 140 patients included in the study, 51 developed VS after treatment, representing an incidence rate of 36.4%. The SE-Inception model yielded superior performance (accuracy: 0.93; area under the receiver operating characteristic curve [AUC]: 0.95), surpassing both ResNet50 (accuracy: 0.85; AUC: 0.90) and Random Forest (accuracy: 0.59; AUC: 0.65). Recall and F1 scores also improved markedly, indicating robust sensitivity and precision. Calibration curves demonstrated excellent agreement between predicted and observed risks, while decision curve analysis (DCA) consistently indicated superior net clinical benefits of the SE-Inception model across various threshold probabilities compared to ResNet50 and Random Forest. Grad-CAM consistently localized to anatomically relevant regions correlating with surgeon- and radiologist-identified risk sites, strengthening the clinical interpretability and trustworthiness of the predictive framework.

Conclusions: Taking the surgical context into account, our SE-Inception framework demonstrated enhanced accuracy and interpretability in identifying patients at risk for postoperative radiotherapy-induced VS. Through alignment with expert clinical assessments and enabling early, personalized intervention strategies, this approach has the potential to improve outcomes and long-term quality of life in cervical cancer survivors, supporting more proactive, surgery-informed treatment planning.