Deep learning models based on CT predict prognosis in patients with locally advanced esophageal squamous cell carcinoma who received neoadjuvant immunotherapy and chemotherapy.
Authors
Affiliations (5)
Affiliations (5)
- Department of Radiation Oncology, Shandong Cancer Hospital and Institute, Shandong First Medical University and Shandong Academy of Medical Sciences, Jinan, Shandong, China.
- Department of Oncology, Cancer Center, West China Hospital, Sichuan University, Chengdu, Sichuan, China.
- Department of Gastroenterology, Shandong Cancer Hospital and Institute, Shandong First Medical University, Shandong Academy of Medical Science, Jinan, Shandong, China.
- Department of Respiratory Oncology, Shandong Cancer Hospital and Institute, Shandong First Medical University and Shandong Academy of Medical Sciences, Jinan, Shandong, China.
- Division of Thoracic Tumor Multimodality Treatment, Cancer Center, West China Hospital, Sichuan University, Chengdu, Sichuan, China.
Abstract
Neoadjuvant immunotherapy and chemotherapy (NICT) has emerged as the standard pharmacological intervention for locally advanced esophageal squamous cell carcinoma (ESCC). However, individual responses to this drug combination vary significantly. This study aims to develop a deep learning (DL) framework based on spatial radiomics to non-invasively profile tumor responses to NICT, thereby serving as a companion diagnostic to design optimal personalized combinatorial regimens. This retrospective study included 655 patients with locally advanced ESCC receiving NICT across two centers (January 2022-December 2024), divided into training, internal, and external validation cohorts. We utilized VISTA3D combined with point-prompt technology for precise primary tumor segmentation. Unsupervised clustering was then applied to delineate intratumoral heterogeneous "habitats", which capture the microenvironmental phenotypes associated with immunochemotherapeutic response. A DL model based on ResNet18 and an attention-based multi-instance learning (MIL) framework was constructed to extract spatial radiomics features, predict disease-free survival (DFS), and stratify patients to guide the design of adjuvant therapeutic strategies. The AI-assisted segmentation achieved a robust Dice similarity coefficient of 0.87 ± 0.18 in the external validation cohort. The ResNet18-based DL model demonstrated excellent prognostic performance (External C-index:0.85, 95% CI:0.71-0.95). By translating imaging phenotypes into risk profiles, the model successfully stratified patients: the low-risk group exhibited significantly superior DFS and OS compared to the high-risk group (P<0.001). Crucially, the model facilitated precise treatment regimen design. High-risk patients derived substantial survival benefits from the addition of adjuvant radiotherapy (DFS HR:0.52; OS HR:0.82), whereas low-risk patients showed no significant benefit, suggesting they could be spared from radiation-induced toxicities. Grad-CAM and SHAP analyses confirmed that the model's predictions correlate with tumor invasion and microenvironmental heterogeneity. The proposed AI-driven spatial radiomics framework successfully bridges macroscopic imaging phenotypes with pharmacological responses to NICT. As an advanced companion diagnostic tool, it optimizes the design of multi-modality combinatorial regimens, preventing overtreatment while maximizing survival benefits for high-risk ESCC patients in the era of precision oncology.