A total of 535 CC patients (mean age, 51.22 ± 10.89 years, range, 22–76 years) from three centers were included. Of which, the 251 patients (mean age, 50.90 ± 10.77 years, range, 25–71 years) from center A were used as the training set, and the other 284 patients (mean age, 51.51 ± 10.98 years, range, 22–76 years) from center B and C were used as an independent external validation set. The histopathological characteristics are presented in Table 1. No significant differences were found between the training and validation set in clinical characteristics and volumetric ADC histogram metrics (all p > 0.05) (Table 2).

The ICCs showed moderate interobserver agreement for maxADC, 0.70 [95% confidence interval (CI): 0.51–0.82], and excellent agreements for maximum diameter, 0.98 (95% CI: 0.96–0.99); volume, 0.98 (95% CI: 0.94–0.99); skewness, 0.80 (95% CI: 0.68–0.88); kurtosis, 0.81 (95% CI: 0.69–0.89); minADC, 0.92 (95% CI: 0.87–0.95); meanADC, 0.95 (95% CI: 0.89–0.98); P10_ADC, 0.99 (95% CI: 0.99-1.00); P25_ADC, 0.99 (95% CI: 0.98–0.99); P50_ADC, 0.97 (95% CI: 0.94–0.99); P75_ADC, 0.92 (95% CI: 0.79–0.96); P90_ADC, 0.84 (95% CI: 0.69–0.92), respectively. The Bland–Altman plots showed small interobserver variabilities (Fig. 3).

Bland-Altman plots for the interobserver variabilities of MRI-measured diameter and volumetric ADC histogram metrics. The X-axis shows the mean of two observers, and the y-axis shows the difference between them. The continuous blue lines represent the mean bias between the two observers, and the dashed lines represent the 95% limits of agreement of the mean difference

Clinical characteristics and volumetric ADC histogram metrics comparisons between different subtypes, LVSI and LNM status in training set.

The ASC/AC had significantly lower proportions of SCC-Ag > 1.1 ng/ml and higher proportions of CA199 ≥ 37 U/ml and HE4 ≥ 70 pmol/l than the SCC group (all p < 0.05). Volumetric ADC histogram analysis showed there was a decreasing trend in tumor ADC values (except for maxADC) and entropy and an increasing trend in skewness and kurtosis from AC, ASC, to SCC subtypes. Compared with the SCC group, the ASC/AC had significantly higher values in entropy, minADC, meanADC, P10_ADC, P25_ADC, P50_ADC, P75_ADC, and P90_ADC and significantly lower values in skewness and kurtosis than SCC group (all p < 0.05) (Table 3).

Compared with the ASC group, the AC had a significantly lower proportion of SCC-Ag > 1.1 ng/ml and significantly higher values in meanADC, P10_ADC, P25_ADC, P50_ADC, and P75_ADC (all p < 0.05). No significant differences were found in patients’ age, menopause status, serum CA125, CA153, CEA, SF, AFP, tumor diameter, and volume among SCC, ASC, and AC subtypes (all p > 0.05) (Table 3). The comparisons of clinical characteristics and volumetric ADC histogram metrics are shown in Fig. 4.

The clinical and MRI quantitative metrics comparison between cervical squamous cell carcinoma (SCC), adenosquamous carcinoma (ASC), and adenocarcinoma (AC). (a) The ADC histogram of three cases shows the distributions of the ADC values of all voxels within the whole tumor of SCC (blue), ASC (green), and AC (red). (b-r) Comparison of clinical and ADC histogram metrics between SCC and ASC, ASC and AC, and SCC and AC groups.* P < 0.05; ** P < 0.01; *** P < 0.001; ns, no significance

The LVSI(+) CCs had a significantly higher proportion of CEA ≥ 5 ng/ml, larger diameter, volume, and maxADC, and lower entropy than the LVSI(-) group (all p < 0.05). LNM(+) CCs had significantly higher proportions of CA153 ≥ 31.3 U/ml and CA199 ≥ 37 U/ml, larger diameter, volume, and maxADC, and lower entropy than LVSI(-) group (all p < 0.05) (Table 3).

In differentiating the ASC/AC from SCC, the P25_ADC, SCC-Ag, and CA199 were independent predictors and constructed the nomogram model (recorded as nomogram_1, Fig. 5, a). The odds ratios (OR) were 5.95 (95%CI: 3.12–11.34) for P25_ADC, 0.06 (95%CI: 0.02–0.14) for SCC-Ag, and 6.71 (95%CI: 1.93–23.33) for CA199, respectively. Nomogram_1 achieved an AUC of 0.900 (95% CI: 0.851–0.948) in the training set and 0.873 (95% CI: 0.805–0.941) in the validation set, respectively (Fig. 5, b). The sensitivity, specificity, accuracy, PPV, and NPV were 85.5%, 83.2%, 83.7%, 58.8%, and 95.3% for the training set, and 78.9%, 87.4%, 86.3%, 49.2%, and 96.4% for the validation set, as shown in Table 4 and confusion matrix in Fig. 5, c and d. The calibration curves show good agreement between predicted probabilities and actual observations (Fig. 5, e and f). Hosmer-Lemeshow test also showed the Nomogram_1 was well-fitted (P values were 0.589 and 0.229 for the training and validation sets, respectively). The DCA shows good net clinical benefit across a broad range of threshold probabilities in both training and validation sets (Fig. 5, g and h).

Nomogram model for differentiating ASC/AC from SCC. (a) Nomogram_1 is constructed by independent predictors of P25_ADC, SCC-Ag, and CA199. For each patient, points were assigned for each variable, and a total score was calculated from the nomogram. The total points corresponded to a predicted probability of ASC/AC. (b) The receiver operator characteristic curves (ROCs) for nomogram_1 in training set (red line) and validation set (green line). (c, d) The corresponding confusion matrix of training set (c) and validation set (d). (e, f) The calibration curves show that the predicted probabilities are in good agreement with the actual observations. (g, h) The decision curves of training (g) and validation sets (h) show positive net clinical benefits for the nomogram_1 across a broad range of threshold probabilities

Further analysis showed the nomogram_1 had an AUC of 0.960 (95% CI: 0.920–0.999) in differential AC from SCC in the training set, with a sensitivity of 90.6%, a specificity of 94.9%, an accuracy of 94.3%, a PPV of 73.4%, and an NPV of 98.4%; and an AUC of 0.945 (95% CI: 0.887-1.000) in the external validation set, with a sensitivity of 91.3%, a specificity of 94.7%, an accuracy of 94.4%, a PPV of 61.8%, and an NPV of 99.1%.

In differentiating the AC from ASC, the P50_ADC and SCC-Ag were independent predictors and constructed the nomogram model (recorded as nomogram_2, Fig. 6, a). The OR were 5.29 (95% CI: 1.74–16.09) for P50_ADC and 0.11 (95%CI: 0.03–0.45) for SCC-Ag, respectively. Nomogram_2 achieved an AUC of 0.837 (95% CI: 0.731–0.943) in the training set and 0.829 (95% CI: 0.701–0.957) in the validation set, respectively (Fig. 6, b). The sensitivity, specificity, accuracy, PPV, and NPV were 71.9%, 87.0%, 78.2%, 88.5%, and 69.0% for the training set, and 60.9%, 93.3%, 73.7%, 93.3%, and 60.9% for the validation set, as shown in Table 4 and confusion matrix in Fig. 6, c and d. The calibration curves show good agreement between predicted probabilities and actual observations (Fig. 6, e and f). Hosmer-Lemeshow test also showed the Nomogram_2 was well-fitted (P values were 0.118 and 0.922 for the training and validation sets, respectively). and the DCA shows good net clinical benefit across a broad range of threshold probabilities in both training and validation sets (Fig. 6, g and h).

Nomogram model for differentiating AC from ASC. (a) Nomogram_2 is constructed by independent predictors of P50_ADC and SCC-Ag. For each patient, points were assigned for each variable, and a total score was calculated from the nomogram. The total points corresponded to a predicted probability of AC. (b) The ROCs for nomogram_2 in training set (red line) and validation set (green line). (c, d) The corresponding confusion matrix of training set (c) and validation set (d). (e, f) The calibration curves show that the predicted probabilities are in good agreement with the actual observations. (g, h) The decision curves of training (g) and validation sets (h) show positive net clinical benefits for the nomogram_2 across a broad range of threshold probabilities

The tumor volume is the only independent predictor of LVSI (+) and LNM (+) [ORs were 1.025 (95% CI: 1.004–1.046) and 1.031 (95% CI: 1.014–1.048), respectively]. In predicting LVSI (+), the AUC, sensitivity, specificity, accuracy, PPV, and NPV were 0.608 (95% CI: 0.532–0.684 ), 78.2%, 40.3%, 66.5%, 74.7%, and 44.9% in the training set, and 0.553 (95% CI:0.482–0.624), 78.8%, 24.4%, 62.3%, 70.6%, and 33.3% in the validation set, respectively (Table 4). In predicting LNM (+), the AUC, sensitivity, specificity, accuracy, PPV, and NPV were 0.694 (95% CI: 0.624–0.764), 69.7%, 68.6%, 68.9%, 49.1%, and 83.9% in the training set, and 0.656 (95% CI: 0.584–0.728), 68.3%, 63.4%, 64.8%, 43.1%, and 83.1% in the validation set, respectively (Table 4).