2715 - Development of Radiomic Models Incorporating Neural Networks for Early Prognostic Prediction in Patients with Locally Advanced Cervical Cancer

10:45am - 12:00pm PT

Hall F

Screen: 1

POSTER

Presenter(s)

Chang Cai, MD, BS - Ruijin Hospital, Shanghai JiaoTong University School of Medicine,, Shanghai, Shanghai

C. Cai¹, J. Xiao², R. Cai³, J. Chen⁴, and H. Xu⁵; ¹Ruijin Hospital, Shanghai JiaoTong University School of Medicine,, Shanghai, China, ²Shenzhen United Imaging Research Institute of Innovative Medical Equipment, Shenzhen, China, ³Ruijin Hospital, Shanghai Jiao Tong University School of Medicine, Shanghai, Shanghai, China, ⁴Department of Radiation Oncology, Ruijin Hospital, Shanghai Jiaotong University School of Medicine, Shanghai, China, ⁵Rui Jin Hospital, Shanghai Jiao Tong University, School of Medicine, Shanghai, China

Purpose/Objective(s): To develop predictive models utilizing dynamic MRI data for early treatment, aiming for an accurate and comprehensive prediction of 2-year progression-free survival (PFS) and 2-year overall survival (OS) in patients with locally advanced cervical cancer (LACC), by training neural network models and integrating them with machine learning algorithms to collaboratively extract radiomic features.

Materials/Methods: A total of 172 patients with LACC who received definitive radiotherapy between March 2017 and December 2022 in our center were retrospectively enrolled. These patients were randomly divided into a training set (137 cases) and a test set (35 cases) in an 8:2 ratio. Baseline clinical features and outcome indicators, including 2-year PFS and 2-year OS, were recorded during follow-ups. Neural network models were first trained according to the prediction labels. Radiomic features and neural network features were then extracted from pre-treatment and early-treatment MRI using neural networks and Support Vector Machine (SVM), respectively. The least absolute shrinkage and selection operator (LASSO) regression was applied to select the features used for model construction. Subsequently, SVM was employed to build models in the training set using five-fold cross-validation. The models were then evaluated in the test set to explore the application value of incorporating neural network features in radiomic models for predicting the prognosis of LACC patients based on dynamic MRI changes from pre-treatment to early treatment (?MRI).

Results: The median follow-up time for the patients in this study was 3.7 years (IQR: 2.7–4.9 years), with a 2-year PFS of 74.4% and a 2-year OS of 84.3%. Radiomic models, neural network models, neural network-radiomic models, and neural network-radiomic-clinical models were successfully developed using ?MRI and clinical features. In model evaluation, the neural network-radiomic-clinical models exhibited higher AUC values for predicting 2-year PFS and 2-year OS (0.880 and 0.887, respectively) compared to the other models.

Conclusion: Incorporating neural network features based on dynamic MRI changes from pre-treatment to early treatment, along with clinical features, can significantly enhance the predictive performance of machine learning models in prognostic prediction for LACC patients.