Treatment Response Prediction Using a Combination of Hematological Biomarkers and CT Deep Learning Radiomics in Locally Advanced Esophageal Squamous Cell Carcinoma

To assess the complementary value of hematological biomarkers to deep learning-radiomic models for assessing esophageal squamous cell carcinoma (ESCC) pathological complete response (pCR) after neoadjuvant chemoradiotherapy (nCRT), which will help to provide a reference for the following clinical study of esophageal preservation.A total of 157 patients with ESCC were enrolled and divided into a training cohort (n = 111) and a validation cohort (n = 47). Computed tomography (CT) was performed for all patients 2 weeks before and 6 weeks after nCRT. In addition, clinicopathological factors and hematological parameters before nCRT were collected. Deep learning and handcrafted radiomic features were extracted from segmented regions of interest (ROIs) from pretreatment (ROI1) and posttreatment (ROI2) CT, which represented the features of the pre- and posttreatment tumors, respectively. The minimum redundancy maximum relevance (mRMR) and least absolute shrinkage and selection operator (LASSO) algorithms were used for feature selection, and logistic regression (LR) was used as the classifier. The deep learning radiomic nomogram (DLRN) was then developed based on the rad-scores and independent clinicopathological risk factors. The model was assessed using area under the receiver operating characteristic (ROC) curve analysis, calibration curve analysis and decision curve analysis and verified with both 10-fold cross-validation and internal validation using bootstrap resampling with 1000 replicates.Rad-scores were constructed with 8 features, which were finally selected as the most predictive features from ROI 1 and ROI 2. The neutrophil-to-lymphocyte ratio (NLR), lymphocyte-to-monocyte ratio (LMR), albumin and rad-scores were incorporated into the DLRN, which achieved good prediction performance, with areas under the ROC curve of 0.888 (95% CI, 0.824-0.951, p <0.05) for the training cohort and 0.873 (95% CI, 0.772-0.974, p <0.05) for the validation cohort. On the training set, the DLRN achieved an AUC of 0.882 in 10-fold cross-validation, and after internal validation, the area under the ROC curve still reached 0.884. The DLRN performed significantly better than the clinical model and radiomics models (p<0.05). IDI and continuous NRI showed significant improvement for the DLRN when incorporating radiomics features and hematological parameters (IDI = 0.3399, P <0.001; continuous NRI = 1.141, P <0.001; categorical NRI = 0.3836, P <0.001). Calibration curves (p > 0.05) and DCA demonstrated that the DLRN was useful for pCR prediction and produced a greater net benefit than the clinical model and radiomics models.Incorporation of radiomics features and hematological parameters into the DLRN improved pCR prediction after nCRT in ESCC. Enhanced pCR predictability may improve patient selection before surgery, providing clinical application value for the use of active surveillance.