Preoperative CT Image Prediction of Ovarian Cancer Chemotherapy Response Score Based on Vision Transformer

Researchers developed a multimodal deep learning framework integrating Vision Transformer and clinical data, which can predict the response of high-grade serous ovarian cancer patients to neoadjuvant chemotherapy (Chemotherapy Response Score, CRS) using routine preoperative CT images, providing an early and non-invasive assessment tool for clinical decision-making and facilitating precision medicine.

High-grade serous ovarian cancer (HGSOC) is one of the most aggressive types of gynecological malignancies, with significant biological and spatial heterogeneity. Most patients are diagnosed at an advanced stage. For patients unsuitable for immediate surgery, neoadjuvant chemotherapy (NACT) combined with delayed surgery is the standard regimen. The Chemotherapy Response Score (CRS) is a well-validated pathological marker for NACT response, but it can only be obtained postoperatively. Clinicians cannot predict the response when formulating the initial plan, so preoperative prediction of CRS can help optimize treatment strategies.

The research team proposed an innovative 2.5D multimodal framework, whose core components include:

1. Vision Transformer Encoder: Uses pre-trained ViT to extract visual features, captures long-range dependencies via self-attention, and understands tumor spatial distribution;
2. Lesion-Dense Omentum Slice Processing: Focuses on omentum regions rich in lesions (a common metastasis site for HGSOC) to extract features with predictive value;
3. Intermediate Fusion Module: Integrates visual features with clinical variables (age, tumor markers, staging, etc.), with better interaction effects than early/late fusion;
4. 2.5D Architecture: Processes adjacent slices to capture spatial context, avoiding the high computational cost and overfitting risk of pure 3D methods.

The model was validated on two independent datasets:

• Internal Test Set (IEO Cohort): 41 patients, ROC-AUC of 0.95, accuracy of 95%, precision of 80%, strong discriminative ability under the same center's data;
• External Test Set (OV04 Cohort): 70 patients, ROC-AUC of 0.68, accuracy of 67%, precision of 75%. The decline in external performance reflects differences in image acquisition and patient characteristics between centers, suggesting the need for larger multi-center data, domain adaptation techniques, and image standardization. Although the external AUC decreased, it still indicates that the model captures generalized predictive signals.

Research Significance:

1. Early Decision Support: Preoperative CRS prediction helps formulate personalized plans; patients with low response can adjust chemotherapy or explore other options;
2. Non-invasive Assessment: CT prediction is completely non-invasive and can be repeated for dynamic monitoring;
3. Multimodal Value: Combining imaging and clinical data provides a more comprehensive patient profile;
4. Resource Accessibility: CT equipment is widely available, the method is easy to promote, and no dedicated expensive equipment is needed.

Research Limitations and Improvement Directions:

1. Sample Size Limitation: The internal test sample size is small (41 cases), requiring larger-scale multi-center studies;
2. External Generalization Challenge: Cross-center performance decline requires exploration of robust feature representation and domain adaptation strategies;
3. Interpretability Requirement: Need to deeply study the image regions focused on by the model to enhance clinical acceptance;
4. Prospective Validation: Currently using retrospective data, requiring prospective clinical trials to verify clinical utility.

This study is an important progress of AI in precision medicine for gynecological cancers. It combines ViT technology with clinical needs to develop a potential preoperative decision-making tool. Although there is still a distance from clinical application, it lays the foundation for future research. With the expansion of data scale and algorithm optimization, AI-based chemotherapy response prediction is expected to become an indispensable part of the comprehensive treatment of ovarian cancer.