Radiomics analysis of dual-layer spectral detector CT-derived iodine density maps for predicting the pathological complete response of esophageal squamous cell carcinoma after neoadjuvant chemoimmunotherapy Authors: Wanling Mu, Chenglong Luo, Xinhua Meng, Feng Li, Yu Qi, Jinjin Dang, Janbo Gao, Yue Zhou Published: 2026 ABSTRACT PurposeTo evaluate the efficacy of radiomics from dual-layer spectral detector CT (DLCT) iodine density maps, combined with clinicopathological characteristics and spectral parameters, in predicting pathological complete response (pCR) following neoadjuvant chemoimmunotherapy (nCIT) in patients with esophageal squamous cell carcinoma (ESCC).Materials and methodsA total of 129 patients with ESCC undergoing nCIT were allocated to training (n = 86) and validation (n = 43) groups. Patients were categorised into pCR and non-pCR groups based on postoperative tumor regression grade. Multiple parameter maps were reconstructed from pre-treatment DLCT images for quantitative spectral analysis. Radiomic features extracted from the venous-phase iodine density maps were used to develop a radiomics score (Radscore). Univariate and multivariate logistic regression analyses were used to identify independent predictors of pCR, leading to the development of clinical, spectral, radiomics, and combined models. The model performance was evaluated using receiver operating characteristic (ROC) curves, decision curve analysis, and calibration curves.ResultsMultivariate logistic regression analysis identified differentiation grade, venous-phase normalized iodine density (NIDV), and Radscore as independent predictors of pCR. The combined model demonstrated superior predictive performance compared to the individual models, with areas under the curve of 0.943 (training) and 0.833 (validation). Calibration curves showed good agreement between the predicted and observed outcome values, whereas decision curves confirmed a superior net clinical benefit in both cohorts.ConclusionsThe combined model incorporating radiomics features from DLCT iodine density maps, clinicopathological characteristics, and spectral parameters demonstrated robust predictive capability in predicting pCR to nCIT in ESCC, which is anticipated to be a valuable tool for personalized treatment decision-making. Introduction Esophageal cancer (EC) is one of the most common malignant tumors of the digestive system and has high incidence and mortality rates [1]. China has one of the highest prevalence rates of esophageal cancer globally, with approximately 90% of cases being esophageal squamous cell carcinoma (ESCC). Although advancements in surgery, chemoradiotherapy, and targeted therapies have significantly improved patient survival in recent years, the treatment outcomes for advanced and locally advanced esophageal cancer remain suboptimal. The application of immune checkpoint inhibitors (ICIs) offers a promising new approach for enhancing the efficacy of ESCC treatment. Camrelizumab is an anti-PD-1 antibody that enhances the ability of the immune system to fight tumor cells by blocking PD-1 activity [2]. Its use as a neoadjuvant treatment for ESCC is gaining clinical significance and has become a key research area. Studies have shown that neoadjuvant chemoimmunotherapy (nCIT) can shrink the primary tumor, decrease tumor stage, and improve the chances of complete surgical removal without increasing the risk of adverse side effects [3,4]. Pathological complete response (pCR) is widely recognized as a critical indicator for evaluating treatment effectiveness, with studies demonstrating that pCR in the primary tumor is a significant prognostic factor for ESCC [5,6]. However, owing to tumor heterogeneity, a considerable number of patients do not benefit from treatment and may experience disease recurrence, progression, or treatment-related adverse effects. Therefore, early and accurate prediction of nCIT efficacy is essential for timely prognostic assessment and is vital for developing individualized treatment plans. The novel spectral detector CT system utilizes a stereoscopic dual-layer detector. A single conventional scan simultaneously acquires both high- and low-energy data [7]. Spectral-based imaging (SBI) facilitates the generation of diverse image representations and spectral analyses, including iodine density (ID) maps, virtual monochromatic images (VMIs), effective atomic number (Zeff) maps, and electron density (ED) maps. The iodine density reflects the concentration of iodine in tissues and indirectly reflects the blood supply characteristics of tumors [8]. Normalized iodine density (NID) helps minimize variations in individual circulation. Previous studies have investigated the value of iodine-related parameters for predicting tumor treatment responses [9,10]. These studies demonstrated the potential of dual-layer spectral detector CT(DLCT) iodine parameters for treatment response assessment in EC; however, they were limited by small sample sizes, and a lack of integration with radiomics features to capture tumor heterogeneity. Radiomics employs high-throughput extraction of imaging features [11], integrating features reflecting tumor heterogeneity with patient clinical and pathological data into a combined model. This provides a comprehensive characterization of the pathological and physiological characteristics of the tumor, as well as its spatiotemporal heterogeneity. Innovatively, this study utilized spectral CT-derived iodine density maps combined with radiomic features. By integrating relevant clinicopathological information, we established a combined model to non-invasively predict and assess the treatment response to nCIT in ESCC. Materials and methods Patient enrollment The study was conducted in accordance with the Declaration of Helsinki (2013 revision) and was approved by the Ethics Committee of *** (Approval No. 2024-KY-2073-002). The Ethics Committee waived the requirement for individual patient consent. Retrospective data from 129 patients with pathologically confirmed ESCC who underwent nCIT followed by surgical resection at ***between October 2022 and October 2024. Patients from the Heyi campus were assigned to the training cohort (Campus 1, n = 86), whereas those from the East and South campuses were assigned to the validation cohort (Campuses 2 and 3, n = 43). The cohort comprised 96 males and 33 females, aged 44–77 years. The inclusion criteria were as follows: (1) pathologically confirmed ESCC; (2) clinical stage T2–4aN+/-M0 (American Joint Committee on Cancer TNM Staging System, 8th Edition [12]); (3) completion of a contrast-enhanced DLCT scan within two weeks prior to treatment initiation; and (4)undergoing surgical resection after completion of two or three cycles of nCIT (camrelizumab + cisplatin or nedaplatin + albumin-bound paclitaxel). The exclusion criteria were as follows: (1) incomplete clinicopathological data; (2) significant CT artefacts that hinder lesion observation or measurement; (3) prior anti-tumor therapy before the current neoadjuvant regimen; and (4) severe heart, liver, or kidney dysfunction, or other autoimmune diseases. The detailed process for the inclusion and exclusion patients is shown in Fig. 1.Fig. 1Flowchart of patient enrollment. DLCT, dual-layer spectral detector CT; ESCC, esophageal squamous cell carcinoma; nCIT, neoadjuvant chemo-immunotherapy; pCR, pathologic complete response. Flowchart of patient enrollment. DLCT, dual-layer spectral detector CT; ESCC, esophageal squamous cell carcinoma; nCIT, neoadjuvant chemo-immunotherapy; pCR, pathologic complete response. Imaging equipment and acquisition protocol All patients underwent dual-phase (arterial and venous) contrast-enhanced spectral CT scans (Philips IQon Spectral CT, Best, Netherlands) after oral administration of 800–1000 mL of water. Scan coverage extended from the thoracic inlet to the pubic symphysis. Conventional 120-kVp blended images were reconstructed using the iDose algorithm, whereas spectral reconstruction generated the SBI. Subsequent processing was performed on Philips IntelliSpace Portal V9.0 and arterial- and venous- phase parametric maps from the SBI series were derived, including ID maps, Zeff maps, ED maps, and 40,100 keV VMIs. Detailed information regarding the spectral CT scanning protocol and data reconstruction is provided in Supplementary A1. Clinicopathological data assessment Clinicopathological data included age, sex, smoking status, alcohol consumption history, and tumor differentiation grade. The differentiation grade was obtained from the pretreatment endoscopic biopsy pathology reports and categorized as either poorly differentiated or well/moderately differentiated. The detailed neoadjuvant therapy regimens are provided in Supplementary A2. Patients were stratified by tumor regression grade (TRG), which was assessed using surgical specimens according to the criteria established by the College of American Pathologists. TRG classification was defined as follows: TRG 0, no viable tumor cells; TRG 1, single or small foci of residual tumor cells; TRG 2, residual tumor nests with stromal fibrosis; TRG 3, minimal or no tumor regression with extensive residual tumor. Patients with TRG grade 0 were classified into the pCR group and all others were categorized as non-pCR. Conventional CT feature assessment was independently conducted by two radiologists (with 17 and 8 years of diagnostic experience) using mediastinal window settings. The evaluated parameters included tumor location (upper, middle, lower thoracic esophagus), tumor thickness (measured in the sagittal plane), CT attenuation during the arterial and venous phases, and clinical T/N stage. Any initial discrepancies in interpretation were resolved by consensus reading. Spectral parameter quantification The spectral-enhanced arterial- and venous-phase SBI data sets were analyzed using the Spectral CT Viewer module on the Philips IntelliSpace Portal. Regions of interest (ROIs) were delineated at the largest cross-section of the lesions by two radiologists with 8 and 5 years of diagnostic experience. During placement, care was taken to avoid necrotic and cavitary areas, ensuring that each ROI encompassed more than two-thirds of the lesion, with sizes ranging from 20.12 to 206.88 mm². The following parameters were measured: ID, VMI attenuation at 40 keV (VMI40) and 100 keV (VMI100), Zeff, ED, and the following were calculated: normalized iodine density (NID), arterial enhancement fraction (AEF) and spectral curve slope (λ) were as follows: NID = ID lesion / ID aorta (ID aorta); AEF = (ID arterial / ID venous)× 100%; λ= (VMI40 - VMI100) / (100−40). For NID calculation, a circular ROI covering more than two‑thirds of the aortic lumen was placed in the descending aorta at the same axial level as the tumor ROI, carefully avoiding calcifications or thrombi. All measurements were performed in triplicate, and the mean values from the two radiologists were subsequently used for the final analysis. Fig. 2 illustrates the measurement methodology used in this study.Fig. 2Schematic Diagram of Spectral Parameter Measurements. a-e, A pCR patient. a, Conventional venous phase image at maximal tumor level (CT value=77.1 HU); b, Arterial phase iodine density map at same level (tumor IDA=1.63 mg/mL, same-level aorta IDA=14.21 mg/mL, NIDA=0.11); c, Venous phase iodine density map at same level (tumor IDV=1.54 mg/mL, same-level aorta IDV=4.86 mg/mL, NIDV=0.32); d, Venous phase 40-keV virtual monoenergetic image at same level (VMI40=180.6 HU); e, Venous phase effective atomic number map at same level (Zeff=8.25). F-J, a non-pCR patient (TRG grade 2). f, Conventional venous phase image at maximal tumor level (CT value=92.8 HU); g, Arterial phase iodine density map at same level (tumor IDA=1.87 mg/mL; same-level aorta IDA=11.32 mg/mL, NIDA=0.17); h, Venous phase iodine density map at same level (tumor IDV=2.31 mg/mL, same-level aorta IDV=5.71 mg/mL, NIDV=0.40); i, Venous phase 40-keV virtual monoenergetic image at same level (VMI40=232.9 HU); j, Venous phase effective atomic number map at same level (Zeff=8.34). Schematic Diagram of Spectral Parameter Measurements. a-e, A pCR patient. a, Conventional venous phase image at maximal tumor level (CT value=77.1 HU); b, Arterial phase iodine density map at same level (tumor IDA=1.63 mg/mL, same-level aorta IDA=14.21 mg/mL, NIDA=0.11); c, Venous phase iodine density map at same level (tumor IDV=1.54 mg/mL, same-level aorta IDV=4.86 mg/mL, NIDV=0.32); d, Venous phase 40-keV virtual monoenergetic image at same level (VMI40=180.6 HU); e, Venous phase effective atomic number map at same level (Zeff=8.25). F-J, a non-pCR patient (TRG grade 2). f, Conventional venous phase image at maximal tumor level (CT value=92.8 HU); g, Arterial phase iodine density map at same level (tumor IDA=1.87 mg/mL; same-level aorta IDA=11.32 mg/mL, NIDA=0.17); h, Venous phase iodine density map at same level (tumor IDV=2.31 mg/mL, same-level aorta IDV=5.71 mg/mL, NIDV=0.40); i, Venous phase 40-keV virtual monoenergetic image at same level (VMI40=232.9 HU); j, Venous phase effective atomic number map at same level (Zeff=8.34). Radiomics analysis workflow The radiomics analysis pipeline consist of tumor segmentation, feature extraction, and feature selection. Venous-phase iodine density maps were uploaded to the Deepwise Multimodal Research Platform (v2.3.1, Beijing). Two blinded radiologists (8 and 5 years of experience) independently delineated the ROIs across the maximal tumor cross-section and three adjacent slices, while avoiding necrotic areas and luminal structures. To ensure feature reproducibility, a senior radiologist (with 17 years of experience) randomly selected 30 patients from the training cohort for ROI redrawing. Features with an interobserver intraclass correlation coefficient (ICC) < 0.75 were excluded. The platform extracted seven feature classes per ROI, and the seven categories of radiomics features are fully specified in Supplementary A3. Feature extraction and selection were performed on the training cohort, and all extracted features underwent Z-score normalization. Highly correlated radiomic features (Spearman’s r > 0.9) were excluded to mitigate multicollinearity. Subsequently, the optimal radiomic features were selected using the Least Absolute Shrinkage and Selection Operator (LASSO) regression algorithm, and the Radscore was calculated. Model construction and performance evaluation Variables demonstrating statistically significant differences in the univariate analysis within the training cohort were incorporated into the multivariate logistic regression analysis to identify the independent predictors of pCR. A radiomics model was developed based on the Radscore, and clinical and spectral models were constructed using the identified independent predictors of pCR. A combined model integrating these parameters was subsequently established, and a nomogram was generated for visual representation. ROC curve analysis was performed to evaluate the predictive efficacy of each model for nCIT response and to validate it in the validation cohort. The DeLong test was used to compare the areas under the curve (AUC) across the models. Calibration curves were plotted to assess the model calibration performance. Decision curve analysis (DCA) was conducted to quantify the clinical utility of each model. The complete profile of the study cohort is shown in Fig. 3.Fig. 3Study design and flowchart. Study design and flowchart. Statistical analysis Statistical analyses were conducted using the SPSS software (version 25.0) and MedCalc (version 20.0.15). Continuous variables with a normal distribution were compared using the independent-samples t-test and are presented as mean ± standard deviation (mean ± SD). Non-normally distributed variables are presented as medians (25th and 75th percentiles) and were analyzed using the Mann-Whitney U test. Categorical variables were compared using the chi-squared test or Fisher's exact test. Model development and performance evaluation were performed using the R software (version 4.4.0). The Net Reclassification Improvement (NRI) was calculated to quantify the improvement in risk reclassification of the combined model compared with each single predictor model. A positive NRI indicated improved reclassification performance. Statistical significance was set at p-value < 0.05 for all tests. Results Comparison of clinicopathological characteristics Among the 129 patients, 45 (34.9%) achieved pCR, and 84 (65.1%) were classified in the non-pCR group. No significant differences were observed between the two cohorts in terms of age, sex, tumor location, differentiation grade, clinical T/N stage, smoking status, alcohol consumption, or tumor thickness on CT (all p > 0.05, see Supplementary Table A1). Within the training cohort, statistically significant differences emerged between the groups for differentiation grade, clinical N stage, and tumor location (Table 1; p < 0.05). No significant differences were observed in the remaining parameters (p > 0.05). Multivariable logistic regression analysis identified the differentiation grade as an independent predictor of pCR following nCIT (odds ratio [OR] = 0.174, 95% confidence interval [CI]: 0.033–0.910; p = 0.038).Table 1Comparison of clinicopathological characteristics between the two cohorts.CharacteristicsTraining cohort (n = 86)Validation cohort(n = 43)pCR(n = 31)non-pCR(n = 55)pvaluepCR(n = 14)non-pCR(n = 29)p valueAge (years)62.97 ± 7.9663.11 ± 7.390.93465.21 ± 6.1364.93 ± 7.990.908Gender0.3860.179male2142924female101355Tumor location0.0300.862upper thoracic2912middle thoracic2630921lower thoracic31646Differentiation0.0060.045poorly differentiated423214well or moderately differentiated27321215Clinical T-stage0.6901.000271147322378174a2725Clinical N-stage0.0130.0380–12122992–31033520Smoking status0.6330.739former/current1326614never1829815Alcohol consumption0.8970.324former/current1017515never2138914Tumor thickness(mm)35.40(31.20,44.95)41.80(34.30,49.90)0.09538.45(30.88,53.08)41.00(34.10,48.00)0.400Radscore0.66(0.46.0.69)0.23(0.16,0.41)<0.0010.43(0.25.0.58)0.34(0.24.0.44)0.029pCR, pathologic complete response Table 1Comparison of clinicopathological characteristics between the two cohorts.CharacteristicsTraining cohort (n = 86)Validation cohort(n = 43)pCR(n = 31)non-pCR(n = 55)pvaluepCR(n = 14)non-pCR(n = 29)p valueAge (years)62.97 ± 7.9663.11 ± 7.390.93465.21 ± 6.1364.93 ± 7.990.908Gender0.3860.179male2142924female101355Tumor location0.0300.862upper thoracic2912middle thoracic2630921lower thoracic31646Differentiation0.0060.045poorly differentiated423214well or moderately differentiated27321215Clinical T-stage0.6901.000271147322378174a2725Clinical N-stage0.0130.0380–12122992–31033520Smoking status0.6330.739former/current1326614never1829815Alcohol consumption0.8970.324former/current1017515never2138914Tumor thickness(mm)35.40(31.20,44.95)41.80(34.30,49.90)0.09538.45(30.88,53.08)41.00(34.10,48.00)0.400Radscore0.66(0.46.0.69)0.23(0.16,0.41)<0.0010.43(0.25.0.58)0.34(0.24.0.44)0.029pCR, pathologic complete response Comparison of clinicopathological characteristics between the two cohorts. pCR, pathologic complete response Comparison of spectral parameters In the training cohort, significant differences between groups were observed in IDV, λV, NIDA, and NIDV (Table 2, p < 0.05). These parameters were significantly lower in the pCR group than those in the non-pCR group. Multivariate logistic regression identified NIDV as an independent predictor of pCR following nCIT (OR = 0.806, 95% CI: 0.707–0.920, p = 0.001). The spectral parameters of the validation cohort are listed in Table 3.Table 2Comparison of spectral parameters in the training cohort.ParameterspCR (n = 31)non-pCR(n = 55)t/Z valuep valueCT value A77.50(68.50,84.50)82.30(69.50,90.30)-1.336a0.182CT value V90.60(81.45,99.85)92.75(87.20,101.75)-0.729a0.466IDA1.52(1.26,1.65)1.65(1.27,1.96)-1.552a0.121IDV1.75(1.60,2.05)1.94(1.76,2.23)-2.222a0.026NIDA0.12 ± 0.030.15 ± 0.042.945b0.004NIDV0.35 ± 0.060.42 ± 0.083.884b<0.001ZeffA8.17 ± 0.248.25 ± 0.241.363b0.177ZeffV8.26(8.16,8.40)8.32(8.25,8.47)-1.485a0.138ED%A104.40(103.95,104.80)104.60(104.10,105.15)-1.148a0.251ED%V105.00(104.50,105.30)104.70(104.30,105.10)-1.230a0.219VMI40 A167.20(153.30,183.05)184.20(147.85,222.55)-1.408a0.159VMI40 V197.00(178.20,218.05)213.80(194.85,230.80)-1.794a0.073VMI100 A58.80(52.20,63.85)60.60(55.70,66.80)-1.273a0.203VMI100 V66.52 ± 10.6767.03 ± 11.070.205b0.838λA1.91 ± 0.592.08 ± 0.611.231b0.222λV2.15(1.93,2.55)2.41(2.18,2.77)-2.042a0.041AEF0.82 ± 0.160.83 ± 0.210.232b0.817pCR, pathologic complete response; λ, slope of the spectral curve (40–100 keV); A, arterial phase; AEF, arterial enhancement fraction; a, Z-value; b, t-value; CT, computed tomography; ED%, electron density; ID, iodine density; NID, normalized iodine density; V, venous phase; VMI40 and VMI100 CT values at 40 keV and 100 keV virtual monoenergetic images; Zeff, effective atomic numberTable 3Comparison of spectral parameters in the validation cohort.ParameterspCR (n = 14)non-pCR(n = 29)t/Z valuep valueCT value A75.41 ± 12.8283.41 ± 14.371.767a0.085CT value V80.01 ± 8.2391.04 ± 14.852.583a0.013IDA1.18(1.03,1.73)1.73(1.48,1.97)-2.437b0.015IDV1.56 ± 0.331.87 ± 0.303.048a0.004NIDA0.12(0.10,0.14)0.15(0.13,0.17)-3.499b<0.001NIDV0.30 ± 0.060.37 ± 0.063.358a0.002ZeffA8.06 ± 0.198.23 ± 0.192.630a0.012ZeffV8.17 ± 0.168.29 ± 0.182.120a0.040ED%A104.26 ± 0.63104.51 ± 0.940.888a0.380ED%V104.31 ± 0.59104.68 ± 1.041.227a0.227VMI40 A160.36 ± 33.00188.32 ± 37.342.384a0.022VMI40 V178.29 ± 25.45201.27 ± 36.972.092a0.043VMI100 A58.77 ± 9.6862.20 ± 10.571.023a0.312VMI100 V60.31 ± 8.7164.55 ± 12.511.139a0.261λA1.48(1.34,2.15)2.15(1.76,2.44)-2.320b0.020λV2.08(1.70,2.26)2.36(1.88,2.61)-2.216b0.027AEF0.86 ± 0.150.91 ± 0.151.201a0.237pCR, pathologic complete response; λ, slope of the spectral curve (40–100 keV); A, arterial phase; AEF, arterial enhancement fraction; a, t-value; b, Z-value; CT, computed tomography; ED%, electron density; ID, iodine density; NID, normalized iodine density; V, venous phase; VMI40 and VMI100 CT values at 40 keV and 100 keV virtual monoenergetic images; Zeff, effective atomic number Table 2Comparison of spectral parameters in the training cohort.ParameterspCR (n = 31)non-pCR(n = 55)t/Z valuep valueCT value A77.50(68.50,84.50)82.30(69.50,90.30)-1.336a0.182CT value V90.60(81.45,99.85)92.75(87.20,101.75)-0.729a0.466IDA1.52(1.26,1.65)1.65(1.27,1.96)-1.552a0.121IDV1.75(1.60,2.05)1.94(1.76,2.23)-2.222a0.026NIDA0.12 ± 0.030.15 ± 0.042.945b0.004NIDV0.35 ± 0.060.42 ± 0.083.884b<0.001ZeffA8.17 ± 0.248.25 ± 0.241.363b0.177ZeffV8.26(8.16,8.40)8.32(8.25,8.47)-1.485a0.138ED%A104.40(103.95,104.80)104.60(104.10,105.15)-1.148a0.251ED%V105.00(104.50,105.30)104.70(104.30,105.10)-1.230a0.219VMI40 A167.20(153.30,183.05)184.20(147.85,222.55)-1.408a0.159VMI40 V197.00(178.20,218.05)213.80(194.85,230.80)-1.794a0.073VMI100 A58.80(52.20,63.85)60.60(55.70,66.80)-1.273a0.203VMI100 V66.52 ± 10.6767.03 ± 11.070.205b0.838λA1.91 ± 0.592.08 ± 0.611.231b0.222λV2.15(1.93,2.55)2.41(2.18,2.77)-2.042a0.041AEF0.82 ± 0.160.83 ± 0.210.232b0.817pCR, pathologic complete response; λ, slope of the spectral curve (40–100 keV); A, arterial phase; AEF, arterial enhancement fraction; a, Z-value; b, t-value; CT, computed tomography; ED%, electron density; ID, iodine density; NID, normalized iodine density; V, venous phase; VMI40 and VMI100 CT values at 40 keV and 100 keV virtual monoenergetic images; Zeff, effective atomic number Comparison of spectral parameters in the training cohort. pCR, pathologic complete response; λ, slope of the spectral curve (40–100 keV); A, arterial phase; AEF, arterial enhancement fraction; a, Z-value; b, t-value; CT, computed tomography; ED%, electron density; ID, iodine density; NID, normalized iodine density; V, venous phase; VMI40 and VMI100 CT values at 40 keV and 100 keV virtual monoenergetic images; Zeff, effective atomic number Table 3Comparison of spectral parameters in the validation cohort.ParameterspCR (n = 14)non-pCR(n = 29)t/Z valuep valueCT value A75.41 ± 12.8283.41 ± 14.371.767a0.085CT value V80.01 ± 8.2391.04 ± 14.852.583a0.013IDA1.18(1.03,1.73)1.73(1.48,1.97)-2.437b0.015IDV1.56 ± 0.331.87 ± 0.303.048a0.004NIDA0.12(0.10,0.14)0.15(0.13,0.17)-3.499b<0.001NIDV0.30 ± 0.060.37 ± 0.063.358a0.002ZeffA8.06 ± 0.198.23 ± 0.192.630a0.012ZeffV8.17 ± 0.168.29 ± 0.182.120a0.040ED%A104.26 ± 0.63104.51 ± 0.940.888a0.380ED%V104.31 ± 0.59104.68 ± 1.041.227a0.227VMI40 A160.36 ± 33.00188.32 ± 37.342.384a0.022VMI40 V178.29 ± 25.45201.27 ± 36.972.092a0.043VMI100 A58.77 ± 9.6862.20 ± 10.571.023a0.312VMI100 V60.31 ± 8.7164.55 ± 12.511.139a0.261λA1.48(1.34,2.15)2.15(1.76,2.44)-2.320b0.020λV2.08(1.70,2.26)2.36(1.88,2.61)-2.216b0.027AEF0.86 ± 0.150.91 ± 0.151.201a0.237pCR, pathologic complete response; λ, slope of the spectral curve (40–100 keV); A, arterial phase; AEF, arterial enhancement fraction; a, t-value; b, Z-value; CT, computed tomography; ED%, electron density; ID, iodine density; NID, normalized iodine density; V, venous phase; VMI40 and VMI100 CT values at 40 keV and 100 keV virtual monoenergetic images; Zeff, effective atomic number Comparison of spectral parameters in the validation cohort. pCR, pathologic complete response; λ, slope of the spectral curve (40–100 keV); A, arterial phase; AEF, arterial enhancement fraction; a, t-value; b, Z-value; CT, computed tomography; ED%, electron density; ID, iodine density; NID, normalized iodine density; V, venous phase; VMI40 and VMI100 CT values at 40 keV and 100 keV virtual monoenergetic images; Zeff, effective atomic number Radiomic feature selection and interpretation A total of 2153 radiomic features were initially extracted from the venous-phase iodine density maps. Subsequently, 10 optimal features were selected using ICC and LASSO regression to calculate the Radscore (details in Supplementary Figure A1 and A2). Among these, the largest positive contribution came from lbp‑2D_firstorder_10Percentile, a first‑order feature reflecting local texture homogeneity, while the largest negative contribution came from wavelet‑HLH_glcm_SumAverage, a GLCM‑based feature indicating overall enhancement intensity. The GLCM features constituted the largest proportion (40%), suggesting that textural patterns and intensity distribution were the primary drivers of the Radscore (Supplementary Table A2). In both the training and validation cohorts, the Radscore was significantly higher in the pCR group than in the non-pCR group. Multivariate logistic regression analysis confirmed that the Radscore was an independent predictor of pCR following nCIT (OR=1.165, 95% CI: 1.090–1.245, p < 0.001). Violin plots illustrating the distribution of the Radscore in both cohorts are provided in Supplementary Figure A3. Model construction and performance evaluation In the training cohort, the clinical model (differentiation grade), spectral model (NIDV), and radiomics model (Radscore) achieved AUC values of 0.645, 0.741, and 0.869, respectively. In contrast, the combined model (integrating all three) achieved a significantly higher AUC of 0.943, accuracy of 0.884, sensitivity of 0.839, and specificity of 0.946. DeLong's test demonstrated that the combined model had a significantly superior predictive performance compared to each individual model (all p < 0.05). In the validation cohort, the combined model yielded an AUC of 0.833. The ROC curves for all models in both cohorts, along with the results of DeLong's test, are shown in Fig. 4. A nomogram visualizing the combined model is shown in Fig. 5. Calibration curves indicated a strong agreement between the predicted probabilities and actual outcomes for the combined model in both cohorts. DCA revealed that the combined model provided a greater net clinical benefit than the individual models across clinically relevant threshold probabilities in both cohorts (Fig. 6). Specifically, the combined model outperformed all three single models across threshold ranges of 1%–99% in the training cohort and 14%–93% in the validation cohort. The detailed performance metrics for all models are provided in Table 4. To further quantify the reclassification performance, we calculated the NRI, as shown in Table 5; all of which were positive and statistically significant (all p < 0.05).Fig. 4ROC comparison of various models. a, Training and b, validation cohorts. c, DeLong tests for the training cohort revealed that the combined models outperformed to all individual models (all P < 0.05). Clinic clinical, Rad radiomics, ROC receiver-operating characteristic.Fig. 5The nomogram of combined model. NIDV normalized iodine density in the venous phase.Fig. 6Decision curves in the cohorts, a, training cohort, b, validation cohort. c, the calibration curves of combined models in training and validation cohort, demonstrated a close agreement between model predictions and actual observations.Table 4Performance comparison between the models.CohortModelAUC(95%CI)AccuracySensitivitySpecificityTraining cohortClinical0.645(0.534–0.745)0.6400.8710.418Spectral0.741(0.636–0.830)0.6510.6770.781Radiomics0.869(0.780–0.932)0.8020.8070.818Combined0.943(0.871–0.982)0.8840.8390.946Validation cohortClinical0.653(0.492–0.791)0.6740.8570.448Spectral0.778(0.626–0.891)0.7210.7140.724Radiomics0.685(0.525–0.818)0.7670.7140.690Combined0.833(0.687–0.929)0.7910.6430.931AUC, Area under the curve; CI, confidence intervalTable 5NRI of the combined model compared with single predictor models.ComparisonTraining cohortValidation cohortNRI (95% CI)p valueNRI (95% CI)p valueCombined vs. Clinical1.479 (1.185–1.749)< 0.0010.734 (0.132–1.328)0.019Combined vs. Spectral1.459 (1.110–1.759)< 0.0010.670 (0.077–1.240)0.028Combined vs. Radiomics1.039 (0.680–1.397)< 0.0010.882 (0.299–1.428)0.002NRI values > 0 indicate improved reclassification performance of the combined model over the corresponding single-variable model. 95% confidence intervals were estimated via bootstrap with 1000 resamples. ROC comparison of various models. a, Training and b, validation cohorts. c, DeLong tests for the training cohort revealed that the combined models outperformed to all individual models (all P < 0.05). Clinic clinical, Rad radiomics, ROC receiver-operating characteristic. The nomogram of combined model. NIDV normalized iodine density in the venous phase. Decision curves in the cohorts, a, training cohort, b, validation cohort. c, the calibration curves of combined models in training and validation cohort, demonstrated a close agreement between model predictions and actual observations. Table 4Performance comparison between the models.CohortModelAUC(95%CI)AccuracySensitivitySpecificityTraining cohortClinical0.645(0.534–0.745)0.6400.8710.418Spectral0.741(0.636–0.830)0.6510.6770.781Radiomics0.869(0.780–0.932)0.8020.8070.818Combined0.943(0.871–0.982)0.8840.8390.946Validation cohortClinical0.653(0.492–0.791)0.6740.8570.448Spectral0.778(0.626–0.891)0.7210.7140.724Radiomics0.685(0.525–0.818)0.7670.7140.690Combined0.833(0.687–0.929)0.7910.6430.931AUC, Area under the curve; CI, confidence interval Performance comparison between the models. AUC, Area under the curve; CI, confidence interval Table 5NRI of the combined model compared with single predictor models.ComparisonTraining cohortValidation cohortNRI (95% CI)p valueNRI (95% CI)p valueCombined vs. Clinical1.479 (1.185–1.749)< 0.0010.734 (0.132–1.328)0.019Combined vs. Spectral1.459 (1.110–1.759)< 0.0010.670 (0.077–1.240)0.028Combined vs. Radiomics1.039 (0.680–1.397)< 0.0010.882 (0.299–1.428)0.002NRI values > 0 indicate improved reclassification performance of the combined model over the corresponding single-variable model. 95% confidence intervals were estimated via bootstrap with 1000 resamples. NRI of the combined model compared with single predictor models. NRI values > 0 indicate improved reclassification performance of the combined model over the corresponding single-variable model. 95% confidence intervals were estimated via bootstrap with 1000 resamples. Discussion Recent studies have demonstrated that neoadjuvant chemoimmunotherapy significantly improves pCR rates compared with neoadjuvant chemotherapy alone in patients with locally advanced ESCC, correlating with enhanced disease-free survival and overall survival [13,14]. In the present study, 45 patients (34.9%) achieved pCR, which is consistent with previous reports [15,16]. Since pCR serves not only as a key efficacy endpoint but also as a critical prognostic factor in ESCC, it is associated with prolonged overall survival and improved survival outcomes [17]. This study presents the first integration of spectral multiparametric imaging with radiomics to develop and validate a noninvasive model for predicting pCR. The combined model achieved an AUC of 0.943 in the training cohort. This approach provides quantifiable imaging biomarkers for tumor regression, which may help guide personalized therapeutic strategies for ESCC patients undergoing neoadjuvant therapy. Multivariable logistic regression analysis identified differentiation grade as an independent predictor of pCR following nCIT. We observed that patients with well- or moderately differentiated tumors achieved pCR more frequently than those with poorly differentiated tumors. The biological rationale lies in the inverse correlation between the differentiation status and tumor aggressiveness. Tumors with higher differentiation exhibit histological features that are closer to those of normal cells, possess lower malignant potential, and demonstrate slower proliferation rates. Consequently, these tumors show greater sensitivity to chemotherapy and a higher propensity to achieve a pCR. This finding aligns with a retrospective analysis of EC patients receiving neoadjuvant chemoradiotherapy, which revealed that well-differentiated tumors had significantly higher pCR rates [18]. Feng [19] identified the differentiation grade as a significant factor influencing the efficacy of neoadjuvant chemotherapy in advanced gastric cancer. However, relying solely on the differentiation grade provides limited specificity for predicting pCR, highlighting the necessity for the integration of multimodal biomarkers. Iodine density maps illustrate voxel-wise distributions of iodine concentrations, facilitating the quantitative assessment of lesion enhancement. These maps offer superior characterization of tumor vascularity and microenvironment compared with conventional CT scans [20], providing a novel quantitative biomarker for treatment response evaluation. Venous-phase iodine density maps were chosen because of their more pronounced iodine density changes compared to the arterial phase. This is likely due to the greater stability in contrast agent uptake and retention within lesions, reduced susceptibility to haemodynamic variations [21], and a typically stronger enhancement of EC during the venous phase. This methodological choice is further supported by recent evidence demonstrating that venous‑phase imaging improves image reproducibility, reduces artifacts, and has been successfully used in prior spectral CT and radiomics studies on gastrointestinal cancers [[22], [23], [24]]. Significant intergroup differences were observed in the spectral parameters IDV, λV, NIDA, and NIDV. Subsequent multivariable analysis identified NIDV as the sole independent predictor of pCR. Consistent with previous research findings [[24], [25], [26]], the NIDA and NIDV values were significantly lower in the pCR group than in the non-pCR group, reflecting reduced iodine avidity in treatment‑responsive tumors. However, reliance on a single spectral parameter (NIDV) yielded only moderate predictive performance (AUC 0.741), whereas recent studies have demonstrated that integrating multiple DECT parameters such as extracellular volume fraction with arterial enhancement fraction (AUC 0.893) [26] or venous‑phase NID with effective atomic number (AUCs 0.825 and 0.794) [24] improves predictive accuracy. This evidence supports our multimodal approach combining spectral, clinical, and radiomic features. We hypothesize that higher-grade tumors exhibit increased proliferation and neovascularization, potentially manifesting as enhanced iodine avidity ("iodine-hungry" phenotype). However, the spectral model that incorporated only NID parameters demonstrated a suboptimal AUC (0.741), with limited accuracy (0.651) and sensitivity (0.677), indicating that isolated spectral parameters inadequately predicted the treatment response. Numerous studies have utilized radiomics to predict pathological responses following EC treatment [[27], [28], [29], [30]], consistently demonstrating that radiomic models outperform conventional CT-based assessments in terms of diagnostic performance. The integration of radiomics with spectral CT presents a novel approach for predicting response to neoadjuvant therapy. For example, Liu [31] developed a nomogram that combined venous-phase iodine density maps with radiomics to achieve high diagnostic efficacy in differentiating benign from malignant thyroid nodules. Similarly, Feng [22] employed radiomic analysis of spectral CT iodine density maps to preoperatively predict tumor deposits in colorectal cancer, reporting a significantly higher diagnostic accuracy than conventional CT. This advancement will aid in personalized treatment planning. In this study, the radiomics model based on venous-phase iodine density maps achieved an AUC of 0.869 in the training cohort, underscoring its strong predictive potential for the nCIT response. However, the AUC decreased to 0.685 in the validation cohort, indicating relatively low stability and limited generalizability. Shi et al.'s study on 2D CT radiomics for neoadjuvant immunotherapy plus chemoradiotherapy also demonstrated a similar performance trend, with training and test AUCs of 0.89 and 0.80, respectively [32]. Nevertheless, further integration of the clinicopathological and spectral parameters yielded a combined model with substantially improved performance, with an AUC of 0.943 in the training cohort and 0.833 in the validation cohort. By leveraging the synergistic effects of imaging features and clinical parameters, the combined model offers a more precise, robust, and comprehensive assessment to inform clinical decision-making. This study has several limitations that need to be acknowledged. First, as this was a retrospective study with a limited sample size, selection bias may have persisted despite the inclusion of a multi-campus validation cohort to evaluate the robustness of the model. Therefore, future multicentre, large-scale prospective validation studies are warranted. Second, the radiomics model used in this study exhibited a certain degree of limited generalizability, which may be attributable to the high feature‑to‑event ratio and relatively small validation sample. Future studies are required to validate its generalizability in independent settings. Third, the delineation of ROIs carries inherent subjectivity due to interobserver variability, and our 2.5D approach does not fully capture three‑dimensional tumor heterogeneity. Future studies may explore fully automated 3D volumetric segmentation techniques to improve reproducibility and comprehensiveness. Finally, as PD-L1 expression testing was not available for every patient, it was not included. This association with PD-L1 should be further explored in future studies. Conclusion In summary, our innovative multimodal framework integrating spectral CT-derived iodine density maps radiomics with clinicopathological and spectral parameters demonstrated robust predictive efficacy for pCR following nCIT in patients with ESCC. This approach has significant potential for enhancing the precision of clinical decision-making. Nevertheless, our findings require further validation in larger, prospective, multicentre studies. Abbreviations λ Slope of the curve, AEF Arterial enhancement fraction, AUC Area under the curve, CI Confidence interval, CT Computed tomography, DCA Decision curve analysis, DLCT dual-layer spectral detector CT, EC esophageal cancer, ED electron density, ESCC esophageal squamous cell carcinoma, ID Iodine density, ICC Intraclass correlation coefficient, LASSO Least Absolute Shrinkage and Selection Operator, nCIT neoadjuvant chemo-immunotherapy, NID normalized iodine density, OR Odds Ratio, pCR pathologic complete response, ROC Receiver operating characteristic, ROI Region of interest, SBI Spectral-based imaging, TRG tumor regression grade, VMI Virtual Monochromatic Images, Zeff Effective atomic number Ethics declaration Written informed consent to take part in the study and to publish the article has been obtained from all participants or their legal representatives. The privacy rights of participants have been observed. This study was performed in compliance with relevant laws, regulatory frameworks and guidelines where the research took place. This study was conducted in accordance with Declaration of Helsinki (2013 revision). This study was approved by the the Ethics Committee of The First Affiliated Hospital of Zhengzhou University. (Approval No. No.2024-KY-2073-002). Ethics statement The study was conducted in accordance with the Declaration of Helsinki and was approved by the Ethics Committee of The First Affiliated Hospital of Zhengzhou University (Approval No.2024-KY-2073–002). The ethics committee granted a waiver for individual patient consent due to the retrospective nature of this study. Funding This study was supported by the Science and Technology Research Project of Henan Province (No. 252102310073), and Zhengzhou Basic Research and applied Basic research Project (No. 2024ZZJCYJ007). Declaration of Competing Interest The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.