Multimodal Al for Precision Diagnosis of Esophageal Cancer

This study intends to construct two multimodal deep learning models: one for the diagnosis of esophageal cancer and the prediction of invasive depth to assess suitability for endoscopic resection; the other model, based on this, classifies endoscopic non-resectable patients into different degrees of invasion to further explore the differences in the sensitivity and survival of AI-predicted benign and malignant tumors in patients' responses to NAT, thereby providing reliable decision support for precise individualized treatment. This aspect has rarely been addressed in previous studies.

Inclusion Criteria: * Age ≥18 years. * Histologically confirmed or clinically suspected esophageal squamous cell carcinoma (ESCC). * Availability of pre-treatment endoscopic images and contrast-enhanced chest/upper abdominal CT scans. * Availability of complete clinical and pathological data. * Patients who underwent endoscopic resection (ESD/EMR) or esophagectomy with pathological assessment of tumor invasion depth. * Adequate image quality for analysis. * Written informed consent (for prospective cohorts, if applicable). Exclusion Criteria: * Histology other than squamous cell carcinoma. * Prior treatment for esophageal cancer before baseline imaging, including chemotherapy, radiotherapy, immunotherapy, or endoscopic resection. * Distant metastatic disease at diagnosis. * Incomplete clinical, imaging, or pathological data. * Poor-quality CT or endoscopic images unsuitable for analysis. * History of another active malignancy within the past 5 years. * Recurrent esophageal cancer.