Fati, Francesca

Abstract Ovarian cancer is the deadliest gynecologic malignancy worldwide, with a 5-year overall survival rate of approximately 49%. Although complete resection is associated with the most favorable prognosis following primary debulking surgery, accurately assessing tumor resectability at diagnosis remains a major clinical challenge. We propose a decision support system (DSS) designed to predict residual tumor after primary debulking surgery, based on clinical and imaging data available at diagnosis. The system was developed and validated using a retrospective cohort of 465 patients with high-grade serous ovarian cancer, collected at the European Institute of Oncology in Milan, Italy. We developed a deep learning (DL) model that combines pre-trained Vision Transformer encoders with an attention mechanism and a classification head. When evaluated on an independent test cohort of 75 cases, our best-performing model achieved an area under the curve (AUC) of 0.80 (95% CI: 0.68–0.89) and a recall of 0.86 (95% CI: 0.71–0.97), demonstrating discriminatory ability with a particularly low rate of false negatives. Clinically, the model correctly identified 24 out of 28 patients (86%) with non-resectable disease, who would not have benefited from primary debulking surgery. Notably, within this subgroup, the model accurately predicted 93% (13 out of 14) of cases in which surgery was aborted intraoperatively due to unforeseen unresectable disease. These findings suggest that the model could have potential in preventing unnecessary and inappropriate surgical interventions. The proposed DSS is a fully automated, DL–based system for predicting tumor resectability at diagnosis, without the need for manual segmentation or expert evaluation of radiological features and structured clinical parameters. This approach could facilitate and accelerate personalized treatment planning in ovarian cancer. The work is part of the project Under-XAI: understanding ovarian cancer initiation and progression through explainable AI. Project code: PNRR-MAD-2022-12376574.

e13701 Background: The optimal therapeutic strategy for advanced epithelial ovarian cancer (EOC) — primary debulking surgery (PDS) followed by adjuvant chemotherapy vs. neoadjuvant chemotherapy (NACT) followed by interval debulking surgery — remains debated. NACT is often chosen for unresectable disease, typically identified only during surgery. This study aims to develop a deep learning (DL) model to predict tumor resectability from pre-operative computed tomography (CT) scans. Methods: We retrospectively included EOC patients, FIGO stage III-IV, who underwent PDS between 01/ 2016 and 12/2023 at the European Institute of Oncology, Milan, Italy. The dataset included anonymized portal venous phase contrast-enhanced CT scans. Residual tumor (RT) annotations were extracted from operative reports and categorized as optimal RT (< 1cm) and suboptimal RT (≥1cm). A DL segmentation model (based on open-source TotalSegmentator) was used to identify pelvic/abdominal organ volumes and to remove couch artifacts. The dataset was split into training/validation (80%) and testing (20%) sets. A pre-trained Vision Transformer (ViT) encoder was adapted for 3D imaging, converting CT volumes into features refined by an attention mechanism and classified as optimal RT or suboptimal RT . We trained two ViT configurations (vit-base-patch16-224 and dinov2-small-imagenet1k-1-layer) and compared them with DenseNet121, a widely used Convolutional Neural Network (CNN) for image classification. Performance metrics included accuracy and F1 score. Results: Out of 521 patients, 416 (80%) had optimal RT and 105 (20%) suboptimal RT , the proportion being respected througouth the training/validation and testing groups. The median age at diagnosis was 60.0 years (IQR, 53.0-67.0). ViT models showed better performance compared to CNN in terms of accuracy (0.80 and 0.77 vs. 0.62) and F1 score (0.43 and 0.33 vs. 0.12). Conclusions: A pioneering ViT-based model was capable to predict tumor resectability from pre-operative CT scans in patients with advanced EOC. This improvement in performance, compared to CNN, might be due to the ViTs' ability to capture long-range dependencies and global context, which are crucial for predicting tumor resectability. In contrast, CNNs are limited by localized feature extraction and inductive bias, hindering their ability to capture the complex spatial relationships. By addressing the limitations of current radiological assessments, the model has the potential to serve as a clinical decision-support tool, enabling personalized management strategies for patients. Future work will focus on further improving the model performance.