Predicting Pathological Complete Response in Esophageal Squamous Cell Carcinoma Using a Multimodal Model Integrating Clinical, Radiomics, and Deep Learning Features

Design and Setting. Multicenter, retrospective cohort study conducted at three affiliated hospitals in China. A total of 363 consecutive ESCC patients met eligibility criteria and were split into a training cohort (n=107), internal validation cohort (n=45), and two external test cohorts (n=129 and n=82).

Population. Inclusion criteria: biopsy-confirmed ESCC; locally advanced disease by AJCC 8th edition (cT1N1-T3N0-3M0) on contrast-enhanced CT; completion of standardized neoadjuvant chemo-immunotherapy; availability of high-quality venous-phase chest CT (slice thickness ≤5 mm) within 14 days before therapy; R0 resection 6-8 weeks post-treatment; and a definitive postoperative pathology report documenting pCR. Key exclusions: non-squamous histology, distant metastasis, synchronous malignancies, poor/no venous-phase imaging, slice thickness >5 mm, severe artifacts, incomplete tumor visualization, incomplete treatment, or missing endpoints.

Neoadjuvant Regimen and Imaging. Patients generally received two cycles of taxane-platinum chemotherapy (paclitaxel plus cisplatin or carboplatin) combined with camrelizumab every 2-3 weeks prior to surgery. CT imaging was standardized to venous-phase contrast with 1-5 mm slices; scans without venous phase or >5 mm thickness were excluded. Tumor volumes were delineated by two radiologists; disagreements were adjudicated by a senior radiologist, and features were harmonized via resampling and intensity normalization.

Feature Extraction and Modeling. The pipeline integrated: (1) clinical variables; (2) conventional CT radiomics features (shape, first-order, GLCM, GLRLM, GLSZM, etc.); and (3) 2.5D deep learning slice embeddings aggregated to the patient level using multiple-instance learning (MIL). The 2.5D approach uses adjacent slices in axial/sagittal/coronal planes with ResNet backbones; attention-based MIL plus histogram/BoW-TF-IDF descriptors summarized slice-level predictions. Feature selection used univariate filters, correlation screening, mRMR, and LASSO before training classifiers (logistic regression, SVM, Random Forest, Extra-Trees, LightGBM).

Outcomes and Analysis.

Primary outcome: pCR at surgery (yes/no).

Secondary outcomes: model performance (AUC, sensitivity, specificity, PPV/NPV, calibration) and clinical utility by decision-curve analysis; disease-free survival by Kaplan-Meier analysis.

Ethics. Approved by the Ethics Committee of Nanjing Medical University; informed consent was waived given the retrospective design and use of de-identified data.

Inclusion criteria

Biopsy-confirmed esophageal squamous cell carcinoma (ESCC). Locally advanced disease per AJCC 8th ed. (cT1N1-T3N0-3M0) on contrast-enhanced CT.

Completed standardized neoadjuvant chemo-immunotherapy (e.g., paclitaxel + cisplatin/carboplatin with camrelizumab every 2-3 weeks) prior to surgery.

High-quality venous-phase chest CT (slice thickness ≤5 mm) obtained within 14 days before therapy start.

Underwent R0 resection 6-8 weeks after therapy. Availability of a definitive postoperative pathology report to ascertain pCR status.

Exclusion criteria

Non-squamous histology; distant metastasis (M1); synchronous malignancies. Inadequate imaging quality (no venous phase, slice thickness >5 mm, severe artifacts, or incomplete tumor visualization).

Did not complete the full treatment course or had missing endpoints (e.g., no pathological response record or lost to follow-up).