Deep Learning for Esophageal Cancer Screening Using Volumetric CT: Handling Missing Contrast Modalities Via Knowledge Distillation In a Multi-Center Study
Abstract
Purpose
Esophageal cancer is a lethal malignancy where early detection significantly improves survival. While opportunistic screening using chest CT is promising, the frequent absence of Contrast-Enhanced CT (CECT) in routine exams limits diagnostic accuracy, as Non-Contrast CT (NCCT) lacks soft-tissue differentiation. To develop a multimodal deep learning framework that integrates local and global features for esophageal classification (normal, benign, malignant) and employs a Knowledge Distillation (KD) strategy to ensure robust performance when CECT is unavailable.
Methods
This retrospective, multi-center study included 2,947 patients (Center 1: 1,498; Center 2: 1,449). A hybrid 3D DenseNet-121 + Vision Transformer (ViT) architecture was designed to extract hierarchical features. A multimodal fusion network combined NCCT and CECT information. To address missing modalities, a teacher-student KD framework was implemented, allowing a student model (input: NCCT) to learn "privileged" vascular information from a multimodal teacher. Performance was evaluated using Area Under the Curve (AUC) and Accuracy, with stress tests performed under simulated 100% CECT missingness.
Results
In external testing (Center 2), the proposed DenseNet-121+ViT backbone achieved an AUC of 0.815, outperforming standard CNNs (ResNet-50 AUC: 0.752). The Multimodal Fusion model achieved superior discrimination with an AUC of 0.902 and Sensitivity of 83.2%, significantly surpassing single-modality NCCT (AUC: 0.795) and CECT (AUC: 0.862) models. Under the scenario of 100% missing CECT, the KD-Enhanced Student model maintained robust performance (AUC: 0.812), significantly outperforming the standard NCCT-only model (AUC: 0.795) and the direct padding approach (AUC: 0.735).
Conclusion
The proposed hybrid multimodal framework offers a highly accurate solution for esophageal cancer screening. Crucially, the integrated knowledge distillation strategy effectively compensates for missing contrast scans, enabling the model to deliver "contrast-like" diagnostic performance using only non-contrast images, thereby facilitating scalable opportunistic screening in routine clinical practice.