A Vision Transformer-Radiomics Approach for Enhanced Chemotherapy Outcome Prediction in Ovarian Cancer

Neman Abdoli¹Patrik Gilley¹Ke Zhang¹Youkabed Sadri¹Theresa Thai²Yong Chen²Lauren Dockery²Kathleen Moore²Robert Mannel²Yuchen Qiu^1*

• ¹The University of Oklahoma, Norman, United States
• ²The University of Oklahoma Health Sciences, Oklahoma City, United States

Early prediction of chemotherapy response in ovarian cancer patients is essential for enabling personalized treatment strategies and improving clinical outcomes. However, this prediction remains challenging due to the high heterogeneity of tumor biology, patient-specific factors, and treatment regimens. Recent advances in imaging biomarkers derived from both radiomics and advanced deep learning methods offer promising tools for characterizing tumor phenotypes and predicting treatment outcomes. In this study, we present a predictive methodology that integrates handcrafted radiomics features and transformer-based embeddings. The deep features come from two models: a pretrained vision transformer (ViT) and MedSAM, a foundation model adapted for medical image segmentation. We retrospectively collected CT scans from 182 ovarian cancer patients acquired prior to chemotherapy. From these images, we extracted three groups of features: radiomics descriptors, ViT-based embeddings, and MedSAM-based embeddings. These features were standardized and then processed using the least absolute shrinkage and selection operator (LASSO) regression for optimal feature selection. Support vector machines (SVMs) were used to train classification models for predicting 6-month progression-free survival (PFS). Our results show that the model combining ViT and MedSAM image embeddings achieved the highest AUC of 0.924 ± 0.032. The model integrating all three feature sets attained a comparable AUC of 0.924 ± 0.037 but reached the highest classification accuracy of 0.831 ± 0.042. These findings highlight the effectiveness of combining single-modality multi-method imaging biomarkers and demonstrate the potential of advanced models to generate rich, task-specific tumor representations from CT scans in support of precision oncology.