Development and Validation of a Multimodal Deep Learning Model for Early Esophageal Squamous Neoplasia Detection and Invasion Depth Prediction.
Authors
Affiliations (11)
Affiliations (11)
- Department of Gastroenterology, Changhai Hospital, Shanghai, China.
- Department of Gastroenterology and Hepatology, West China Hospital of Sichuan University, Chengdu, China.
- Sichuan University-University of Oxford Huaxi Joint Centre for Gastrointestinal Cancer, West China Hospital of Sichuan University, Chengdu, China.
- Endoscopy, National Cancer Center,Cancer Institute and Hospital, Chinese Academy of Medical Sciences & Peking Union Medical College, Beijing, China.
- Department of Gastroenterology, Shandong Provincial Hospital Affiliated to Shandong First Medical University, Jinan, China.
- Department of Digestive, First Affiliated Hospital of Shihezi University School of Medicine, Shihezi, China.
- Department of Gastroenterology, Shandong Provincial Hospital affiliated to Shandong First Medical University, Jinan, China.
- Cancer Centre, First Affiliated Hospital of Henan Science & Technology University, Luoyang, China.
- Gastroenterology Department, Sichuan University West China Hospital, Chengdu, China.
- Department of Gastroenterology, Changhai Hospital, Second Military Medical University, Shanghai, China.
- Digestive Diseases, Changhai Hospital, Shanghai, China.
Abstract
Early detection of esophageal squamous cell carcinoma (ESCC) is critical for optimizing patient outcomes. Magnifying endoscopy (ME) and endoscopic ultrasonography (EUS) serve as established diagnostic modalities. MUMA-EDx (Multimodal Ultrasound & Magnifying-endoscopic Algorithm for Early ESCC Diagnostics) integrates deep learning-based ME and EUS imaging to improve early-stage ESCC identification and invasion depth assessment. Model development and internal validation utilized the retrospective dataset, while the prospective cohort served for external validation. MUMA-EDx developed two TResNet_m-based classifiers (ME/EUS) followed by feature-level fusion. Model performance was evaluated using area under the receiver operating characteristic curve (AUC-ROC), accuracy, sensitivity, specificity, positive predictive value, and negative predictive value. MUMA-EDx was developed and validated using a retrospective dataset comprising 460 patients (20,889 images) and subsequently tested prospectively on an independent cohort of 131 patients (9,124 images). The feature-level multimodal approach significantly outperformed single-modality models. For tumor discrimination, the model achieved an AUC of 0.94 (95% CI: 0.92-0.96) in retrospective validation and a perfect patient-level AUC of 1.00 (95% CI: 1.00-1.00) in prospective testing. For the more complex task of multiclass invasion depth classification, it achieved a retrospective AUC of 0.95 (95% CI: 0.88-0.99), which remained strong at 0.80 (95% CI: 0.67-0.87) in the prospective cohort. In a comparative study on invasion depth classification, MUMA-EDx's performance exceeded that of novice endoscopists and was comparable to expert-level diagnostics. MUMA-EDx demonstrably delivers exceptional early ESCC detection and robust invasion depth classification, achieving performance comparable to expert endoscopists and poised to significantly enhance diagnostic precision and patient outcomes.