S1062
Clinical – Upper GI
ESTRO 2026
dynamically increased with each treatment stage (pre- nCRT: 54% → post-nCRT: 61% → pre-surgery: 88%), while the non-pCR probability showed a corresponding decrease. Furthermore, the CoT reasoning provided clear explanations for the predictions, linking specific tumor changes (e.g., "reduced tumor volume and soft tissue density after nCRT") to a high probability of pCR, effectively addressing the "black box" problem of traditional models. KM survival analysis confirmed that patients predicted to achieve pCR had a significantly longer recurrence-free survival (RFS) than those predicted to be non-pCR (P<0.05) Conclusion: Our study introduces an interpretable and accurate model that can dynamically track tumor evolution and precisely predict pCR status in ESCC patients receiving nCRT. By providing a reliable, non-invasive tool for preoperative pCR assessment, this model has the potential to guide personalized clinical decisions. References: no reference Keywords: Esophageal cancer, pCR, Ai model Digital Poster 183 A Multicenter Study on Predicting pCR to Neoadjuvant Chemoradiotherapy in Esophageal Cancer by Fusing CT Imaging and Histopathology QIFENG WANG 1 , HAILIN YUE 1 , LEI WU 1 , LIN PENG 1 , WENCHENG ZHANG 2 , WEI HUANG 3 , Lina Zhao 4 1 Radiation Oncology, Sichuan Cancer Hospital, Chengdu, China. 2 Radiation Oncology, Tianjin Cancer Hospital, Tianjin, China. 3 Radiation Oncology, Shangdong Cancer Hospital, Jinan, China. 4 Radiation Oncology, Xijin Hospital, XiAn, China Purpose/Objective: Pathological complete response (pCR) following neoadjuvant chemoradiotherapy (nCRT) is a key prognostic indicator for survival in patients with esophageal cancer. However, the efficacy of using a single modality, such as macroscopic CT images or microscopic whole slide images (WSIs), to predict pCR is limited. This study aims to build a more accurate multimodal pCR prediction model by integrating macroscopic CT features with microscopic WSI features, thereby enhancing the reliability of clinical assessments. Material/Methods: Data Source and Screening: This was a multicenter study conducted across five hospitals. We obtained a final cohort of 73 patients from Sichuan (internal validation) and 1,100 patients from other centers for external validation (383 from Shandong, 324 from Tianjin, 331 from Xi'an, and 61 from Beijing).We
Digital Poster Highlight 182 A Study on Preoperative pCR Prediction in Esophageal Cancer Patients Using Longitudinal CT Images and Chain-of-Thought Prompt Learning QIFENG WANG 1 , HAILIN YUE 1 , LEI WU 1 , WENCHENG ZHANG 2 , WEI HUANG 3 , LINA ZHAO 4 1 Radiation Oncology, Sichuan Cancer Hospital, Chengdu, China. 2 Radiation Oncology, Tianjin Cancer Hospital, Tianjin, China. 3 Radiation Oncology, Shandong Cancer Hospital, Jinan, China. 4 Radiation Oncology, Xijing Hospital, Xi An, China Purpose/Objective: Esophageal squamous cell carcinoma (ESCC) is associated with high mortality in Asia. While neoadjuvant chemoradiotherapy (nCRT) followed by surgery improves patient survival, the clinical complete response (cCR) is an unreliable predictor of pathologic complete response (pCR), a key factor for long-term survival. Traditional single-timepoint imaging studies fail to capture the dynamic tumor changes during treatment and lack interpretability. There is an urgent need for a more effective and explainable prediction method. Material/Methods: This was a multi-center study conducted across centers from 2021 to 2025. After applying rigorous exclusion criteria for patients who did not receive nCRT, had non-squamous cell carcinoma, abnormal radiation doses, double primary cancers, or missing three-stage CT data, we obtained a final cohort of 290 effective samples for internal validation (Sichuan) and 188 samples for external validation (Beijing: 50, Shandong: 65, Tianjin: 73). We propose a novel longitudinal prediction model based on a Chain-of- Thought (CoT) framework. The core of the model is a "pre-trained encoder - conditional network - CoT prompt" architecture. It first extracts features from three-stage CT images (pre-nCRT, post-nCRT, and pre- surgery) using a pre-trained Vision-Language Model (VLM) encoder. Next, a "Thought-conditional prompt learning" module generates a phased reasoning chain ("thought1-thought2-thought3") to fuse features from each stage. Finally, a conditional network outputs the pCR/non-pCR prediction. Model performance was evaluated using ACC, sensitivity, specificity, AUC, and Kaplan-Meier survival analysis. We also leveraged CT image text descriptions (e.g., "reduced wall thickness, decreased irregularity") to enhance model interpretability. Results: The model demonstrated strong and stable performance in multi-center validation, with AUC values consistently above 0.70 in the external cohorts (e.g., Shandong and Tianjin). In the internal validation cohort (Sichuan), the pCR prediction probability
Made with FlippingBook - Share PDF online