CT-based LD-3DViT framework for prognostic stratification in resectable locally advanced esophageal squamous cell carcinoma: an exploratory analysis of treatment-associated outcomes across risk subgroups

Neoadjuvant therapy followed by surgery is a standard strategy for resectable locally advanced esophageal squamous cell carcinoma (ESCC), but heterogeneous outcomes make selecting between neoadjuvant intervention and upfront surgery challenging. This study developed a deep learning framework that links automated tumor segmentation with prognostic risk estimation by combining a latent diffusion model (LDM) and a three-dimensional Vision Transformer (3D ViT), termed LD-3DViT. The framework extracts deep volumetric features from pretreatment contrast-enhanced computed tomography (CT) to predict survival and identify prognostic strata with differential treatment-associated outcomes. Pretreatment contrast-enhanced CT images from patients with resectable locally advanced ESCC treated between January 2020 and December 2024 were retrospectively analyzed. Patients were randomly divided into training and validation sets at a ratio of 7:3. LD-3DViT integrated LDM-based automated segmentation and denoised representation learning with 3D ViT-based prediction of 3-year disease-free survival (DFS) and prognostic risk index generation. Segmentation was evaluated using Dice similarity coefficient, Jaccard index, and 95% Hausdorff distance; prediction was assessed using area under the curve (AUC) and C-index. A total of 706 patients were included, including 366 who received neoadjuvant therapy followed by surgery and 340 who underwent upfront surgery. The training and validation sets included 494 and 212 patients, respectively. In the validation set, the segmentation module achieved a Dice similarity coefficient of 0.828 (95% confidence interval [CI], 0.795–0.861). For 3-year DFS prediction, LD-3DViT achieved an AUC of 0.838 (95% CI, 0.762–0.914) and a C-index of 0.812 (95% CI, 0.745–0.879). In the low-risk subgroup, no statistically significant difference in 3-year DFS was observed between the neoadjuvant therapy and upfront surgery groups (79.2% vs. 77.1%, P = 0.740); however, this observational comparison was inconclusive. In the high-risk subgroup, neoadjuvant therapy was associated with higher observed 3-year DFS (59.7% vs. 29.6%, P < 0.001) and overall survival (69.4% vs. 44.4%, P = 0.004). LD-3DViT achieved favorable segmentation and prognostic prediction performance and identified prognostic risk strata with differential treatment-associated outcomes in this single-center internally validated cohort. These hypothesis-generating findings warrant prospective multicenter external validation before any clinical translation. Not applicable.