Eligible patients were aged 18–75 years of either sex, with an Eastern Cooperative Oncology Group performance status of 0–1. All cases were pathologically confirmed as esophageal squamous cell carcinoma, and the patients had radiologically and clinically resectable locally advanced disease that was treated with NAT. Only patients with complete clinicopathological and follow-up data were included in the analyses.

Patients were excluded if they had non-squamous histological subtypes of EC, severe systemic comorbidities that contraindicated standard treatment, or synchronous malignancies in other organs.

This study was conducted in accordance with the principles of the Declaration of Helsinki. Ethical approval and a waiver of informed consent were obtained from the Ethics Committee of Zhongshan Hospital Affiliated to Xiamen University.

Histopathological examination was conducted by two experienced pathologists who analyzed the hematoxylin and eosin (H&E)-stained slides to evaluate the resection margins and lymph nodes. Tumor staging was based on the 8th edition of the American Joint Committee on Cancer staging system (26). The tumor response to NAT was assessed using the College of American Pathologists (CAP) grading system (modified Ryan scheme) (8). The specific category-to-criteria mapping is defined as follows: TRG 0 (no viable cancer cells); TRG 1 [single cells or rare small groups of cancer cells, <10% viable residual tumor cells (VRTCs)]; TRG 2 (residual cancer cells outgrown by fibrosis, 10–50% VRTCs); and TRG 3 (minimal or no tumor regression, >50% VRTCs). Patients were categorized as good (TRG 0–1) or poor (TRG 2–3) responders. Overall survival (OS) was defined as the time from diagnosis to death, and progression-free survival (PFS) was defined as the time to tumor progression or death from any cause.

All hematoxylin and eosin (H&E)-stained sections were digitally scanned and subjected to quality control. The TRG score of the patients was determined based on postoperative pathological assessment criteria and served as the primary source of the labels.

This study included 157 patients with EC who underwent NAT followed by surgical resection. H&E-stained WSIs were retrospectively collected, yielding 1,298 digital pathology slides. Patient-level TRG labels were obtained from pathology reports, and the final TRG score was determined by integrating multiple slides per patient. Given that TRG is inherently a case-level construct and patient-level labels may not represent the histology of individual slides (e.g., a slide with no tumor cells from a TRG 2 patient), direct inheritance of labels is imprecise. Therefore, to enable the evaluation of slide-level feature extraction capabilities, an independent dataset of 120 WSIs was randomly sampled. Two experienced pathologists manually assigned a “slide-level TRG” to each WSI solely based on the proportion of residual tumor area observed on that specific slide, following the CAP criteria. These slide-specific annotations served as the ground truth for the technical benchmarking of foundation models.

To extract discriminative features from the WSIs, we systematically evaluated three recently developed pathology foundation models: Prov-Gigapath, TITAN, and CHIEF, using 120 WSIs with slide-level TRG annotations.

Prov-Gigapath is a large-scale model trained on 170,000 WSIs. It integrates a tile encoder with a LongNet-based slide encoder to process gigapixel images. While it offers robustness through scale, its architecture is primarily patch-centric. TITAN, in contrast, adopts a multimodal framework (ConCH v1.5 encoder) designed to integrate images and language. Trained on WSIs paired with captions, it excels in cross-modal alignment and zero-shot predictions, focusing on semantic generalization rather than pure histological hierarchy.

CHIEF, however, introduces a hierarchical tokenization mechanism that allows the model to simultaneously capture fine-grained cellular morphology and broader tissue architecture. This architectural characteristic formed our primary a priori rationale for selecting CHIEF as the optimal backbone. Accurate TRG grading requires the simultaneous recognition of residual tumor cells (local features) and the surrounding fibrotic stroma (global context). CHIEF’s ability to integrate these multi-scale features makes it theoretically more suitable for TRG assessment compared to the patch-centric approach of Prov-Gigapath or the language-aligned focus of TITAN.

Consequently, CHIEF was selected as the designated feature extractor, with the empirical benchmarking serving primarily to validate this theoretical hypothesis. The pre-trained backbone generates a 768-dimensional representation for each slide. For all experiments, features were extracted using these foundation models at 20× resolution and subsequently subjected to downstream analysis.

For each WSI, features were extracted using all three foundation models at 20× resolution and subsequently underwent further analysis.

To evaluate whether the learned embeddings reflected TRG-specific differences, we used nonlinear dimensionality reduction methods, namely t-distributed stochastic neighbor embedding and uniform manifold approximation and projection. These techniques project high-dimensional embeddings onto a two-dimensional space, enabling the qualitative visualization of class separation while preserving local neighborhood relationships and global data structures.

A quantitative evaluation of slide-level classification was performed using support vector machines trained on embeddings from each foundation model. The support vector machine classifier was chosen for its proven effectiveness in high-dimensional feature spaces and robustness against overfitting. A radial basis function kernel was used to capture the nonlinear class boundaries, and the hyperparameters were optimized within each training fold. To ensure robust estimation of model generalization, five-fold cross-validation was applied, with training and testing sets partitioned at the slide level. The performance metrics included overall accuracy, area under the receiver operating characteristic curve (AUC), precision, recall, F₁-score, and confusion matrices.

Following the comparative evaluation of the foundation models at the slide level, the best-performing foundation model was incorporated into a patient-level predictive framework. Because TRG labels were available only at the case level, we adopted an MIL strategy that treats each patient as a bag and the associated WSIs as instances within that bag. The bag inherits the patient-level TRG label, whereas the labels of the individual instances remain unknown. This approach is particularly suitable for weakly supervised pathology tasks, in which detailed annotations are difficult to obtain.

Three MIL variants are implemented and compared. First, attention-based MIL aggregates slide-level embeddings using an attention mechanism that assigns variable importance weights to each instance, highlighting the most informative slides for prediction. Second, adaptive cross-instance MIL (ACMIL) extends this framework by modeling dependencies among slides and dynamically adjusting the weighting scheme to account for inter-instance relationships. Third, the clustering-constrained attention MIL (CLAM) imposes an additional constraint that partitions instances within a bag into clusters representing positive and negative evidence, thereby enhancing the interpretability and robustness of the aggregated representation. In all three cases, CHIEF was used as the feature extraction backbone because it demonstrated the strongest discriminative power in the preliminary experiments.

Patient-level MIL models were trained and evaluated using five-fold cross-validation, with data split at the patient level to avoid information leakage across the folds. The evaluation metrics included overall accuracy, macro-F₁ score, and AUC, with confusion matrices providing detailed insights into class-specific performance.

For each patient, feature embeddings extracted from all available WSIs (ranging from 2 to 10 slides) were aggregated to form a single patient-level bag, enabling the MIL framework to capture intra-patient histological heterogeneity across different tissue sections. No padding or truncation was applied; all slides belonging to a patient were included during bag aggregation. Five-fold cross-validation was performed with patient-level splits to strictly prevent information leakage between training and test sets.

To address class imbalance between favorable (TRG 0–1) and poor (TRG 2–3) response groups, class-balanced resampling was applied at the patient (bag) level during training. Model optimization was conducted using the AdamW optimizer with an initial learning rate of 2 × 10⁻⁴ and a weight decay of 4 × 10⁻⁴. The batch size was set to 2 patients per iteration, and all models were trained for 100 epochs.

A cosine annealing learning rate scheduler with warm-up was employed. The warm-up phase lasted for six epochs with an initial learning rate of 1 × 10⁻⁶. The cosine schedule was initialized with a cycle length of two epochs, with the cycle length doubling after each restart up to a maximum of 20 epochs. The minimum learning rate was set to 1 × 10⁻⁷. To ensure stable optimization under weak supervision, each epoch consisted of a fixed number of 200 training iterations.

The training objective was defined as a label-smoothed cross-entropy loss with a smoothing factor of 0.2, which helped mitigate overconfidence and improve generalization in the presence of noisy patient-level supervision. All hyperparameters were kept identical across cross-validation folds, resulting in negligible variation and ensuring reproducible model training.

Statistical analyses were performed using MATLAB (MathWorks, Natick, MA, United States). Model performance was assessed using accuracy, precision, recall, F₁-score, and AUC with 95% confidence intervals (CIs). Group-wise comparisons were performed using paired t-tests or Wilcoxon signed-rank tests, with significance set at p < 0.05. Uniform manifold approximation and projection were performed for feature visualization and implemented using MATLAB.

Deep learning experiments, including feature extraction from foundation models and MIL frameworks, were conducted in Python (version 3.10) using an NVIDIA GTX 3090 GPU.

We analyzed 157 patients with esophageal squamous cell carcinoma who received NAT treatment. Among them, 126 (80.3%) were male, and 100 (63.7%) were aged ≥60 years. Postoperative pathological outcomes showed that 41 and 116 patients achieved TRG 0–1 and TRG 2–3, respectively. The baseline characteristics of the patients are summarized in Table 1.

We designed a two-stage analytical workflow to systematically evaluate the performance of the pathology foundation models and establish an interpretable AI framework for TRG score prediction (Figure 1). At the slide level, three large-scale pathology foundation models were evaluated using 120 annotated WSIs to extract high-dimensional histopathological embeddings. The best-performing model was integrated into a patient-level prediction framework based on MIL, in which slide-level embeddings served as instances, and patient-level TRG score served as weak supervisory labels.

Furthermore, we projected high-dimensional slide-level embeddings onto 2D using uniform manifold approximation and projection to assess the discriminative capacity of slide-level features extracted with different pathology foundation models (Figure 2). The resulting visualization revealed that the embeddings generated by CHIEF and TITAN demonstrated a clearer separation between the TRG categories, particularly for the extreme classes, such as TRG 0 and TRG 3. In contrast, the features derived from Prov-Gigapath showed substantial overlap across categories, indicating limited discriminability for this specific task. These observations suggest that models pretrained with broader histological diversity (CHIEF) or multimodal objectives (TITAN) were more effective in capturing regression-related morphological variations than the patch-centric Prov-Gigapath model.

The quantitative classification experiments further confirmed these qualitative trends (Table 2). In five-fold cross-validation, CHIEF consistently outperformed the other two models when using support vector machine classifiers trained on slide-level features. It achieved balanced precision, recall, and F₁-scores across most TRG categories, with particularly high performance for TRG 3 (F₁ = 0.933) and acceptable results for intermediate categories such as TRG 1 and TRG 2 (F₁ ranging from 0.556 to 0.746). TITAN achieved competitive performance in TRG 2 and 3 but was less effective in the lower categories. Prov-Gigapath performed poorly overall, with most F₁-scores of 0.40. These findings indicate that CHIEF is the most suitable backbone for downstream patient-level modeling (see Table 3).

Based on the slide-level results, we constructed a patient-level TRG prediction framework using MIL. The dataset comprised 157 patients, each contributing 2–10 WSIs. Because only patient-level TRG labels were available, each patient was treated as a bag containing multiple instances (WSIs), with the CHIEF-extracted features of each slide serving as instance representations.

We implemented and compared three representative MIL approaches, namely attention-based MIL (ABMIL), ACMIL, and CLAM, to aggregate slide-level features into patient-level predictions. The model was trained using a five-fold cross-validation setting to ensure patient-level independence between the folds.

Across the five-fold cross-validation, the models demonstrated robust stability. We report the performance as the mean metrics across the folds. The MIL variants achieved a mean macro-average AUC of 0.810 (95% CI: 0.745–0.865) for ACMIL and 0.820 (95% CI: 0.771–0.880) for CLAM. The mean micro-average AUCs consistently exceeded 0.86. Specifically, the CLAM model achieved the highest mean macro-average AUC, while ACMIL demonstrated the most balanced performance with a mean overall accuracy of 82.7%.

The overall receiver operating characteristic curves (Figure 3) illustrated clear separability across the TRG categories, with TRG 0 and 3 achieving the highest discriminative power (AUCs of 0.955 and 0.859 for CLAM, respectively). Intermediate grades (TRG 1–2) showed moderate but consistent classification performance (AUCs = 0.693–0.771).

The patient-level confusion matrices (Figure 4) revealed that most classification errors occurred between adjacent TRG categories, reflecting the inherent ambiguity in histological response assessment. Across all three MIL models, predictions of TRG 0 and TRG 3 showed high agreement with the ground truth (accuracy >0.85), whereas TRG 1 and TRG 2 were more prone to misclassification owing to overlapping morphological features such as mixed residual tumor and stromal fibrosis.

Notably, the ACMIL model demonstrated the most balanced performance, correctly classifying 84.6% of the TRG 0/1 cases and 80.3% of the TRG 2/3 cases, yielding an overall accuracy of 82.7%. The confusion matrices further revealed that the MIL models captured the ordinal structure of TRG grading, as most mispredictions deviated by only one grade, underscoring the biological continuity between adjacent TRG levels rather than the random errors.

To assess the prognostic relevance of TRG in EC, we examined whether pathologist-assigned or AI-predicted TRG score could stratify patients by long-term survival (Figure 5). Given the clinical convention that treatment responders exhibit lower TRG score, the patients were classified into two groups: TRG 0–1 (favorable response) and TRG 2–3 (poor response). Kaplan–Meier survival curves were generated for PFS and OS, and survival differences were evaluated using the log-rank test. Hazard ratios (HRs) and 95% CIs were estimated using Cox proportional hazard regression.

When pathologist-assigned TRG score was used, the difference in OS between the TRG 0–1 and TRG 2–3 groups was not statistically significant (log-rank p = 0.068; HR = 0.428, 95% CI: 0.205–0.894).

In contrast, the AI-predicted TRG score, particularly those generated using the ACMIL model, demonstrated an improved ability to stratify patient outcomes. For PFS, the predicted TRG 0/1 group showed a sustained advantage over the predicted TRG 2/3 group throughout the follow-up period [HR = 0.312, 95% confidence interval (CI), 0.151–0.642; p = 0.0196]. A consistent pattern was observed in OS. Patients with predicted TRG 0/1 achieved longer OS than those with predicted TRG 2/3 (HR = 0.327, 95% CI: 0.157–0.682, p = 0.0259).

These results suggest that AI-derived TRG score captures the histopathological correlates of treatment response more effectively than conventional grading, providing superior discrimination of long-term outcomes. Specifically, ACMIL-predicted TRG has emerged as a robust surrogate biomarker for patient prognosis after NAT, offering potential value for response assessment and clinical trial endpoint refinement.