Inclusion criteria were: a) fibergastroscopic biopsy-proven esophageal squamous cell carcinoma (SCC) before treatment; b) chest enhanced CT examination at baseline; c) chemotherapy or CRT; d) diagnosis with esophageal fistula by CT, endoscopy, barium esophagography or operation at follow-up; e) availability of quality diagnostic images for measuring lesions. Patients who were not in accord with inclusion criteria were excluded. Patients with esophageal cancer accompanied by esophageal fistula after chemotherapy or CRT were identified from January 2011 to December 2019. In addition, esophageal SCC patients with no esophageal fistula after chemotherapy or CRT were included in this study from January 2016 to December 2017 in a 1:3 ratio. The patients were allocated to the test cohort and validation cohorts in a 1:1 ratio according to the time of inclusion. The complete patient enrollment process is shown in Figure 1.
Chemotherapy or CRT
Patients treated with chemotherapy received a platinum-based regimen, mainly including paclitaxel (175 mg/m2 i.v. on day 1 of every 3-week cycle) and cisplatin (75 mg/m2 i.v. on day 1 of every 3-week cycle). Other patients were treated with concurrent chemoradiotherapy, with a radiotherapy dose of 50-60Gy and platinum-based chemotherapy.
All patients received enhanced multi-slice CT (MDCT) scanning of the chest before and during chemotherapy or CRT. Scans were performed using a 64-row helical CT scanner (Lightspeed VCT; General Electrical Medical Systems, Milwaukee, WI, USA). All patients were in the supine position. Generally, the scan began at 2.0 cm above the lung apices and extended through the adrenal glands. The following imaging parameters were used: 120–140 kVp tube voltage; 300 mA tube current; 64 x 0.625 mm detector collimation; 0.6 s/rotation gantry rotation speed, and 1.5 helical pitch. Axial, coronal and sagittal images were reconstructed using a section width of 5.0 mmOne hundred milliliters of the non-ionic contrast medium iohexol (Omnipaque 300; GE Healthcare) was injected at a rate of 3.0 mL/s through the median cubital vein.
MDCT imaging data were transferred to a picture archiving and communication system (PACS). Two radiologists with 8 years (Dr. Wei) and 12 years (Dr. Shi) of experience in thoracic CT independently reviewed the axial and reconstructed CT images obtained at baseline. Both reviewers were blinded to final results about esophageal fistula. All qualitative and quantitative parameters were assessed on enhanced images before treatment. Final quantitative measurements were determined by averaging the values obtained by the two radiologists. For qualitative analysis, the diagnosis was confirmed by a third experienced radiologist in case of disagreement.
Tumor locations were classified as cervical, upper thoracic, middle thoracic, or lower thoracic esophagus. An important factor in assessing cancer location was to determine the center of the tumor in the esophagus. The tumor was staged by MDCT before therapy according to AJCC/TNM classification, 8th edition. The MDCT status was defined as follows  : CT T0, wall thickness <5.0 mm and no signs of adventitial penetration; CT T1–2, a wall thickness of at least 5–10 mm without evidence of adventitial penetration; CT T3, tumor exhibiting a wall thickness of >10 mm, possibly appearing as ill-defined, abnormal soft tissue around the tumor but no invasion of adjacent structures; CT T4a, invasion of the pleura, pericardium and diaphragm; T4b, invasion of the aorta, vertebral body and trachea. In the last two stages, the tumor had a wall thickness of >10 mm and invaded adjacent structures. Intrathoracic and abdominal lymph nodes >10 mm and supraclavicular lymph nodes >5.0 mm in short-axis diameter were considered metastatic lymph nodes  . N staging was classified as negative (N-) or positive (N+) metastatic lymph nodes. CT imaging findings related to tracheal or bronchial invasion of the tumor were analyzed. The tumor range was classified as four types, including 0-1/4, 1/4-1/2, 1/2-3/4 and 3/4-1 (Figure 2). Esophageal stenosis and deep ulcer were also evaluated. The morphological patterns of the tumor were graded as focal or diffuse.
Tumor wall thickness (THK-tumor) and thinness (THN-tumor) of esophageal SCC were measured perpendicularly to the lumen on axial images using the workstation’s electronic caliper. In case of invisible lumen, the maximal tumor diameter was obtained and multiplied by 0.5. Tumor length (the tumor’s longest diameter, L-tumor) was measured on sagittal CT images. The region of interest (ROI) of the tumor’s maximum CT value (HU-max) was placed on the highly enhanced area and that of the tumor’s minimum CT value (HU-min) on the lowly enhanced area (Figure 3). The area of each ROI was 3-5 mm2, averaging three measurements. The depths of low and high intensity enhancement areas were also measured (Figure 3). Tumor ulceration was quantitatively assessed by measuring ulcer depth (DEP-ulcer), the thickness of the residual wall in the ulcer layer (THK-residue), and the thickness of the lesion adjacent to the ulcer (THK-adjacency) on cross-sectional CT images (Figure 2, 3). In case of no ulcer in the tumor, DEP-ulcer was recorded as 0, and THK-residue and THK-adjacency were the same as the wall thickness of the tumor. The ulcer-to-tumor ratio (R-ulcer) was calculated by the following equation: DEP-ulcer/THK-adjacency. The THK-residue-to-tumor ratio (R-residue) was derived as THK-residue/THK-adjacency. The HU-min to HU-max (R-HU) ratio was obtained as HU-min/HU-max. The joint predictive efficiency of continuous variables was defined as Y1. Y1 combining qualitative signs was defined as Y2.
Definition of esophageal fistula
Esophageal fistula was defined as a connection between the esophagus and adjacent organs or tissues  detected by CT, endoscopy, barium esophagography or operation. On CT images, esophageal fistula was diagnosed by discontinuous or defective esophageal wall, gas and fluid accumulated around the esophagus, or pneumonia associated with esophageal fistula. By esophagography, esophageal fistula was identified as contrast medium leakage into the mediastinum or bronchus.
Differences in qualitative parameters in patients with esophageal SCC between the test and validation cohorts were assessed by the Mann-Whitney test or the Chi-square test/Fisher’s exact test. Differences in quantitative factors were examined by independent-samples t test or the Mann-Whitney test. The associations of quantitative measurements were evaluated by Pearson correlation coefficient; a coefficient > 0.6 suggested a moderate or strong correlation. Only parameters with weak correlation were substituted into the multivariate equation. Receiver operating characteristic (ROC) curve analysis was applied to evaluate the predictive capability of the quantitative analysis for predicting esophageal fistula, with the area under the ROC curve (AUC). Intraclass correlation coefficients (ICCs) were determined to evaluate inter-observer agreement in terms of parameter extraction. Data analysis was conducted with SPSS 22.0 (IBM Corporation, Armonk, NY, USA) and STATA 12.0 (Stata Corporation, College Station, TX, USA).