Development of an improved diagnostic nomogram for preoperative prediction of small cell neuroendocrine cancer of the cervix

Abstract

Objectives:

Accurate preoperative diagnosis of small cell neuroendocrine cancer of the cervix (SCNECC) is crucial for establishing the best treatment plan. This study aimed to develop an improved, non-invasive method for the preoperative diagnosis of SCNECC by integrating clinical, MR morphological, and apparent diffusion coefficient (ADC) information.

Methods:

A total of 105 pathologically confirmed cervical cancer patients (35 SCNECC, 70 non-SCNECC) from multiple centres with complete clinical and MR records were included. Whole lesion histogram analysis of the ADC was performed. Multivariate logistic regression analysis was used to develop diagnostic models based on clinical, morphological, and histogram data. The predictive performance in terms of discrimination, calibration, and clinical usefulness of the different models was assessed. A nomogram for preoperatively discriminating SCNECC was developed from the combined model.

Results:

In preoperative SCNECC diagnosis, the combined model, which had a diagnostic AUC (area under the curve) of 0.937 (95% CI: 0.887–0.987), outperformed the clinical-morphological model, which had an AUC of 0.869 (CI: 0.788–0.949), and the histogram model, which had an AUC of 0.872 (CI: 0.792–0.951). The calibration curve and decision curve analyses suggest that the combined model achieved good fitting and clinical utility.

Conclusions:

Non-invasive preoperative diagnosis of SCNECC can be achieved with high accuracy by integrating clinical, MR morphological, and ADC histogram features. The nomogram derived from the combined model can provide an easy-to-use clinical preoperative diagnostic tool for SCNECC.

Advances in knowledge:

It is clear that the therapeutic strategies for SCNECC are different from those for other pathological types of cervical cancer according to V 1.2021 of the NCCN clinical practice guidelines in oncology for cervical cancer. This research developed an improved, non-invasive method for the preoperative diagnosis of SCNECC by integrating clinical, MR morphological, and apparent diffusion coefficient (ADC) information.

Introduction

Cervical cancer is the fourth most common cancer in females worldwide, with a prevalence of 6.6% and a mortality of 7.7%. ¹ Neuroendocrine carcinomas of the cervix (NECCs) account for 2–5% of cervical malignancies, ^2–5 with significantly lower 5-year overall survival rates of approximately 30% compared to >65% for squamous cell carcinoma and adenocarcinoma of the cervix. ^6,7 As a very aggressive cancer, small cell neuroendocrine carcinoma of the cervix (SCNECC) belongs to the NECC category of high-grade neuroendocrine carcinomas.

The NCCN cervical cancer clinical practice guidelines in oncology suggest different treatment and management strategies for SCNECC than for other subtypes of cervical cancers. ⁸ Specifically, the treatment strategies for SCNECC are based on either locally advanced or metastatic diseases confined to the cervix, whereas those of other subtypes of cervical cancer are based on the staging as specified by the 2018 International Federation of Gynaecology and Obstetrics (FIGO) and whether the treatment is fertility-sparing. ⁸ Notably, SCNECC is not considered suitable for fertility-sparing surgery. ⁸ Cisplatin combined with etoposide chemotherapy is recommended as the first line of chemotherapy for SCNECC, while cisplatin-based chemotherapy alone is recommended for other subtypes of cervical cancer. In addition, brain MR is the recommended imaging modality for SCNECC but not for other subtypes of cervical cancer. ⁸ Therefore, the accurate preoperative diagnosis of SCNECC is crucial for establishing an appropriate, comprehensive treatment plan.

The diagnosis of SCNECC can be made on cervical biopsy. However, this diagnostic approach has many limitations, such as invasiveness, patient compliance, and intra/interobserver variation. Moreover, it may easily lead to an incorrect diagnosis as poorly differentiated cervical cancer due to the limited amount of tissue, and the neuroendocrine component is sometimes only recognized after hysterectomy. ⁹ In contrast, MR is recommended by the European Society of Urogenital Radiology guidelines ¹⁰ as a useful method for cervical cancer diagnosis for initial staging, response monitoring and evaluation of disease recurrence because of its good soft tissue resolution and non-invasiveness. A few reports have offered discussions on the MR characteristics of NECC, but a definitive NECC diagnosis based on preoperative MR results seems difficult. ^11–14 Some studies have reported that low apparent diffusion coefficient (ADC) values are suggestive of NECC. ^12,14 Others suggest that a whole-lesion-based ADC histogram analysis can reflect pathological heterogeneity, which can be achieved by quantifying the ADC intensity distribution in regions of interest (ROIs), which is often ignored by empirical approaches. ^15,16 To date, research on preoperative SCNECC diagnosis has been very limited despite its importance, which may partially be attributed to the rarity of this malignancy. Therefore, this study focused on developing a non-invasive, practical, and sufficiently accurate method for preoperatively diagnosing SCNECC by constructing a combined risk model integrating clinical features, MR morphological features, and whole-lesion ADC histogram analysis.

Methods and materials

Patients

Our Institutional Review Board approved this retrospective study and waived the requirement for informed consent. A total of 69 patients with pathologically (biopsy or surgical resection sampling) confirmed SCNECC from five medical centres, including the First Hospital of Lanzhou University (LZU1^st), Gansu Provincial Maternity and Child-care Hospital (GSMCH), Gansu Tumor Hospital (GSTH), Gansu Provincial Hospital (GSPH), and the Second Hospital of Lanzhou University (LZU2^nd) , between July 2010 and April 2021 were retrospectively reviewed (Supplementary Figure S1). The inclusion criteria included the following: (1) pelvic MR imaging performed before preoperative treatment, (2) MR imaging with good quality can identify the tumour morphology, (3) complete clinical information, and (4) MR performed at 1.5 T and diffusion-weighted (DW) imaging with b values of 50 and 800 s/mm². A patient was excluded if she met any one of the following conditions: (1) missing preoperative pelvic MR imaging, (2) incomplete clinical information, (3) indeterminable boundary and morphology of the SCNECC on MR, or (4) MR performed at 3.0 T. Ultimately, a total of 35 patients with SCNECC who met the inclusion and exclusion criteria were enrolled (Table 1).

Table 1.

Characteristics of patients with cervical cancer

Characteristics	Non-SCNECC	SCNECC	p value
	(N = 70）	(N = 35）
Age (mean ± SD)	48.77 ± 7.61	50.09 ± 10.81	0.472
FIGO staging (n (%))			0.254
I	36 (51.4%)	17 (48.6%)
II	30 (42.9%)	12 (34.3%)
III	2 (2.86%)	4 (11.4%)
IV	2 (2.86%)	2 (5.71%)
Tumour LR Axis (median [IQR])	3.28 [2.66;4.14]	3.30 [2.72;5.18]	0.426
Tumour AP Axis (median [IQR])	2.70 [2.11;3.54]	3.30 [2.08;4.80]	0.16
Tumour height (median [IQR])	3.09 [2.11;3.81]	3.18 [2.12;4.20]	0.681
Tumour mini-diameter (median [IQR])	2.46 [1.76;3.15]	2.82 [2.05;3.78]	0.241
LN long axis (median [IQR])	1.04 [0.84;1.23]	1.36 [0.98;2.16]	0.007
LN short axis (median [IQR])	0.66 [0.53;0.86]	0.98 [0.65;1.85]	0.001
LN tumour ratio (median [IQR])	0.20 [0.14;0.26]	0.27 [0.19;0.51]	0.001
LNM (n (%))			<0.001
Absent	55 (78.6%)	14 (40.0%)
Present	15 (21.4%)	21 (60.0%)
HPV18(n (%))			<0.001
Absent	62 (88.6%)	15 (42.9%)
Present	8 (11.4%)	20 (57.1%)
Tumour size (median [IQR])	3.80 [2.78;4.38]	4.08 [2.80;5.40]	0.274
Tumour size >= 4 cm (n (%))			0.533
Absent	40 (57.1%)	17 (48.6%)
Present	30 (42.9%)	18 (51.4%)
Tumour size >= 5 cm (n (%))			0.002
Absent	62 (88.6%)	21 (60.0%)
Present	8 (11.4%)	14 (40.0%)

AP: anteroposterior; FIGO: International Federation of Gynaecology and Obstetrics; HPV: human papillomavirus; IQR: interquartile range; LNM: lymph node metastasis; LR: left-right;SCNECC: small cell neuroendocrine cancer of the cervix; SD: standard deviation.

A total of 70 non-SCNECC patients were included in this study, including 35 adenocarcinoma (AC) patients and 35 squamous cell carcinoma (SCC) patients (Table 1, Supplementary Figure S1). All patients underwent a pelvic MR imaging examination. The details are shown in Supplementary S1, Supplementary Table S1 and S2.

Histopathology

All patient histopathology diagnoses were confirmed by biopsy or surgical resection sampling findings. Human papillomavirus (HPV) 18 genotyping was performed by polymerase chain reaction (PCR).

Histogram imaging analysis

Anonymous DWI and T ₂-weighted imaging (T2WI) data were archived offline in DICOM format and transferred to FireVoxel workstation (FireVoxel, 301; https://firevoxel.org). The corresponding ADC maps were generated automatically according to the monoexponential diffusion model.

The first reader (Y.S.X, a gynaecologic radiologist with 2 years of experience) manually examined the ROI segments slice by slice on the low b-value DW images (b = 50 s/mm²) along the inside edge of the lesions to delineate the whole tumour volume without excluding cystic, haemorrhagic or necrotic areas for more complete information with reference to the corresponding level of the axial and sagittal T2WI. During the process of ROI segmentation, care was taken to avoid inclusion of any of the adjacent normal cervical stroma or vascular structures and contamination by areas of gas, fluid, or Nabothian cysts in the cervix and artefacts. The second reader (Y.Z.L, a gynaecologic radiologist with 9 years of experience) independently verified the ROIs. Both readers were blinded to the clinical information or the pathological patterns of the patient.

The ADC map within the whole-tumour ROIs was constructed automatically. The histogram and frequency table of ADC values were generated from the ROI. The parameters calculated included the following: minimum, maximum, mean, median, standard deviation (SD), inhomogeneity, skewness, kurtosis, and entropy. The whole-tumour frequency distribution table of the ADC values was imported into SPSS V 26.0 statistical software (IBM, Armonk, NY), and the 5^th, 10^th, 15^th, 20^th, 25^th, 30^th, 35^th, 40^th, 45^th, 50^th, 55^th, 60^th, 65^th, 70^th, 75^th, 80^th, 85^th, 90^th, and 95^th percentiles were subsequently obtained.

Feature selection and model construction

The selected clinical features included age, staging (FIGO 2009 ¹⁷ and FIGO 2018 ¹⁸ ), and HPV18. The selected MR morphological features included lymph node metastasis (LNM), tumour LR (left to right) axis, tumour AP (anteroposterior) axis, tumour height, tumour size, tumour mini-diameter, lymph node long axis, lymph node short axis, and LN-tumour ratio. The definitions and classification criteria of the above features are shown in Supplementary Table 2. Values of the selected features were standardised with the z-score method.

These morphological features were reviewed independently by two radiologists (Y.Z.L and P.L with 9 and 13 years of experience in gynaecologic MR imaging, respectively). Prior to evaluation, the criteria for tumour and lymph node measurements and the determination of lymph node metastasis were standardised for the two radiologists. A variable with p < 0.05 according to Bland–Altman analysis or a κ value less than 0.60 was considered poor and removed.

Histogram features were first assessed by the t test or the Mann-Whitney U-test. Univariate analyses were performed to define the histogram features. The Spearman correlation coefficient was used as an indicator of the correlation among these features, and features with a correlation coefficient of 0.90 or higher were considered redundant and eliminated (Supplementary Figures S3 and S4). All the significant factors were introduced into stepwise logistic regression to select the independent features for the model. Akaike’s information criterion (AIC) was employed as the stopping rule. The retained features were analysed through multivariate logistic regression to build the clinical-morphological model, histogram model, and combined model.

Performance assessment of models

Receiver operating characteristic (ROC) curve analysis was conducted, and the area under the ROC curve (AUC), which describes the discrimination abilities of the model, was obtained as the primary indicator of model performance in SCNECC diagnosis. The Hosmer–Lemeshow test was used to evaluate the goodness of fit of the model. Threefold cross-validation was repeated 10 times to test the diagnostic performance of the models.

Evaluation of clinical applicability

A combined clinical-morphological-histogram nomogram was constructed for assessing its clinical applicability in predicting the probability of SCNECC. Decision curve analysis (DCA) was adopted to assess the clinical usefulness of the models by estimating the net benefits at a range of threshold probabilities. ¹⁹ Three decision curves based on the clinical-morphological model, histogram model, and combined model were plotted.

The entire model-building process is illustrated by the flowchart in Figure 1.

Figure 1.

Feature workflow and study flowchart

Statistical analysis

R software (https://www.rproject.org, V 4.0.4) was used for statistical analyses. To compare the proportional differences between the SCNECC group and the non-SCNECC group, the Mann–Whitney U-test was used for continuous variables, and Pearson’s chi-squared test was used for categorical variables. Differences in AUC values between different models were compared using the DeLong test. An AUC value above 0.75 was considered to be good. ²⁰ A Hosmer–Lemeshow test p < 0.05 was considered to indicate good fitness. Two-tailed p < 0.05 was considered statistically significant.

Results

Patient characteristics

The study flowchart is shown in Figure 1. The demographic characteristics of the patients are summarized in Table 1. The study group consisted of 70 non-SCNECC patients, including 35 adenocarcinoma (AC) patients and 35 squamous cell carcinoma (SCC) patients, with a mean age of 48.77 ± 7.61 years, and 35 SCNECC patients with a mean age of 50.09 ± 1.81 years. SCNECC subtypes included pure type (n = 18), mixed with SCCC (n = 3) and mixed with AC (n = 14). The squamous cell carcinoma type was squamous cell carcinoma (n = 35). The adenocarcinoma subtypes were mucous (n = 32), clear cell (n = 2), and micropapillary (n = 1).

According to the FIGO staging system, the 35 SCNECC patients were classified as stage I (n = 17), II (n = 12), III (n = 4), and IV (n = 2), and the 70 non-SCNECC patients were classified as stage I (n = 36), II (n = 30), III (n = 2), and IV (n = 2). Twenty-nine patients with SCNECC and 62 patients with non-SCNECC underwent radical hysterectomy and retroperitoneal lymph node dissection. According to the preoperative FIGO staging and postoperative risk factors, chemotherapy, pelvic EBRT, or brachytherapy were selectively implemented at the proper time. No significant difference was observed in age, FIGO staging, tumour LR axis, tumour AP axis, tumour height, tumour mini diameter, tumour size, or tumour size >= 4 cm (all p > 0.05).

Clinical-morphological model

The clinical MR morphological features included LNM, tumour LR axis, tumour AP axis, tumour height, tumour size, tumour size >= 4 cm, tumour size >= 5 cm, tumour mini-diameter, lymph node long axis, lymph node short axis, and LN-tumour ratio.

After applying the t-test, Mann–Whitney U-test, or Pearson’s chi-squared test, HPV18, LNM, lymph node long axis, lymph node short axis, tumour size >= 5 cm, and LN-tumour ratio remained as significant clinical MR morphological features (p < 0.05). These significant features were subjected to stepwise logistic regression, and the AIC was employed as the stopping rule; the parameter with the lowest AIC value was selected. HPV18, tumour size >= 5 cm and LN-tumour ratio remained, and variables with p > 0.05 according to Bland–Altman analysis (Supplementary Figure S2) and LNM, with a κ value of 0.76, were retained to build a clinical-morphological model by multivariate logistic regression.

The regression coefficients of LN-tumour ratio, tumour size, and HPV18 were 1.72, 2.50, and 2.22, respectively (p < 0.01) (Table 2). The constructed clinical-morphological model achieved an AUC of 0.869 with a 95% confidence interval (CI) of 0.788–0.949 (Table 3).

Table 2.

Comparison of the models by multivariate logistic regression analysis

Model	Coefficients	OR (95% CI)	p value
Clinical-morphological model
LN-tumour ratio	1.7201	5.585 (1.857–21.434)	0.0059
Tumour size >= 5 cm	2.5006	12.190 (3.509–49.226)	0.0001
HPV18	2.2211	9.217 (2.845–33.143)	0.0003
Histogram model
ADCMinimum	−0.7426	0.476 (0.206–1.050)	0.0701
ADCInhomogeneity	1.5657	4.786 (1.874–14.939)	0.0029
Combined model
LN-tumour ratio	1.2913	3.637 (1.157–15.202)	0.0479
Tumour size >= 5 cm	1.914	6.780 (1.488–35.729)	0.0165
HPV18	2.6191	13.723 (2.778–86.664)	0.0023
ADCMinimum	−1.0156	0.362 (0.118–0.980)	0.0565
ADCInhomogeneity	1.0735	2.926 (0.850–12.387)	0.1141

ADC: Apparent diffusion coefficient; HPV: human papillomavirus; LN: lymph node;OR: odds ratio.

Table 3.

Comparison of the diagnostic performance of the different models

Model	AUC (95% CI)	ACC	SEN	SPE	PPV	NPV	Cut off
Histogram	0.872 (0.792–0.951)	0.867	0.771	0.914	0.818	0.889	0.362279
Clinical-morphological	0.869 (0.788–0.949)	0.857	0.743	0.914	0.812	0.877	0.418755
Combine	0.937 (0.887–0.987)	0.848	0.914	0.814	0.711	0.95	0.229597

ACC: accuracy,AUC: area under the curve, NPV: negative predictive value; PPV: positive predictive value, SEN: sensitivity, SPE: specificity.

Histogram model

First, after application of the Mann-Whitney U-test, ADC_minimum, ADC_mean, ADC_median, ADC_{standard deviation (SD)}, ADC_{inhomogeneity}, ADC_5th, ADC_10th, ADC_15th, and ADC_20th were identified as significant histogram features (p < 0.05) (Supplementary Table S3). Second, the Spearman correlation coefficient was used as an indicator of the correlation among these histogram features, and features with a correlation coefficient of 0.90 or higher were considered redundant and eliminated (Supplementary Figure S3). Subsequently, ADC_minimum, ADC_median, ADC_SD, ADC_{inhomogeneity}, ADC_5th, and ADC_20th remained (Supplementary Figure S3). Third, the ADC histogram percentile features were subjected to collinearity analysis to identify potential independent variable collinearity problems. The analysis shows that the histogram percentile features were collinear (Supplementary Figure 5). Therefore, ADC_20th was removed, and ADC_5th was retained. Fourth, ADC_5th, ADC_minimum, ADC_median, ADC_SD, and ADC_{inhomogeneity} were subjected to stepwise logistic regression, and AIC was employed as the stopping rule. Finally, ADC_minimum and ADC_{inhomogeneity} remained, and a histogram model was built by multivariate logistic regression. For the histogram model, the regression coefficients were −0.7426 for ADC_minimum, which was significant (p < 0.05), and 1.5657 for ADC_{inhomogeneity}, which was highly significant (p < 0.01) (Table 2). The histogram model achieved an AUC of 0.872 with a 95% CI of 0.792–0.951 (Table 3).

Combined model

The combined model was built based on both the clinical-morphological and histogram features (Table 2, Supplementary Figure 6). The correlation among the five independent variables used in the combined model, i.e., LN-tumour ratio, tumour size, HPV18, ADC_minimum, and ADC_{inhomogeneity}, is shown in the chord diagram in Figure 2.

Figure 2.

Chord diagram shows the correlation between three clinical-morphological features (Tumour size, HPV18, LN tumour ratio) and two histogram features (ADC_minimum and ADC_{inhomogeneity}), but there were no significant correlation within the two groups of features. The values around the circle represent the correlation coefficients. Each link represents a significant correlation according to Pearson’s correlation analysis (p < 0.05). The width of the link and the contact area between an arc and a circle indicate the relative strength between two features.

The results of the comparison between the SCNECC group and the non-SCNECC group regarding their values of ADC_minimum, ADC_{inhomogeneity}, and LN-tumour ratio in the combined model are presented in Figure 3. ADC_minimum was significantly lower in the SCNECC group than in the non-SCNECC group, whereas ADC_{inhomogeneity} and LN-tumour ratio were significantly higher in the SCNECC group than in the non-SCNECC group (Figure 3).

Figure 3.

Violin plot of the comparison of features (a) ADC_minimum, (b) ADC_{inhomogeneity} and (c) LN-tumour ratio. There were significant differences in these features between SCNECC and non-SCNECC in the combined model (p < 0.05). The width of the violin plot represents the probability density. The box plots represent the Q3, median, Q1 (the top, middle, and bottom lines of the box, respectively), the 95% CI (the vertical line on the two sides of the box) and outliers (black dots).

The combined model achieved an AUC of 0.937 (95% CI: 0.887–0.987, accuracy: 0.848, sensitivity: 0.914, specificity: 0.814, PPV: 0.711, NPV: 0.950, cut-off value: 0.2295) (Table 3).

Comparison of diagnostic performance

The diagnostic accuracy and AUC of the combined model were better than those of the clinical-morphological model and histogram model (Table 3). The ROC curves of the three models are presented in Figure 4a. According to the DeLong test, the differences between the combined model and the clinical-morphological model (p = 0.0339) and between the combined model and the histogram model (p = 0.0411) were both statistically significant. The difference in the AUC between the clinical-morphological model and histogram model was not significant (p = 0.9527).

Figure 4.

(a) The ROC curves of the clinical-morphological model, histogram model and combined model. (b) The calibration curves of the three models in predicting SCNECC. The different coloured lines represent the performance of the three models, in which a closer fit to the ideal line represents a better prediction. P values of the Hosmer-Lemeshow test are shown in the legend, and p > 0.05 indicates that the model has a good fitting degree. (c) Decision curve analysis for the three models. Across the various threshold probabilities, the combined model curve showed great net benefit.

The calibration plots of Figure 4b show good accordance between the predicted values and the actual outcomes for SCNECC and non-SCNECC for all three models (Figure 4b). The Hosmer-Lemeshow test suggested that there were no statistically significant differences between the predicted and actual values for the clinical-morphological model (p = 0.117), the histogram model (p = 0.148), and the combined model (p = 0.667), which indicates that all three models had good fitness (Figure 4b).

Three-fold cross-validation showed that the clinical-morphological, histogram, and combined models had good performance, with mean AUCs of 0.8507, 0.8455, and 0.8920, mean sensitivities of 0.9011, 0.9334, and 0.8862, and mean specificities of 0.6220, 0.6156, and 0.6763, respectively. (Supplementary Figure 6)

Clinical applicability of the models

Decision curve analysis (DCA) demonstrated that within a threshold probability range of 9–83%, the combined model showed the greatest net benefit over the treat-all and treat-none schemes in predicting SCNECC, although the two other models also showed good benefit over the treat-all and treat-none schemes (Figure 4c). The nomogram was successfully constructed based on the combined model, allowing easy application of the model in the clinic for diagnosing SCNECC (Figure 5).

Figure 5.

The developed nomogram for predicting the probability of SCNECC. By summing the points of each feature and locating the sum on the total points scale, the estimated probability of SCNECC can be determined.

Discussion

The combined model had high diagnostic performance (AUC = 0.937, sensitivity = 0.914, specificity = 0.814), a good fitting degree (HL p = 0.667), and great net benefit results for clinical decision-making. The superior performance of the combined model may be attributed to the combination of clinical-morphological features and ADC histogram features, allowing the inclusion of more information about SCNECC in the model.

In this study, a nomogram was used to diagnostically differentiate SCNECC from non-SCNECC. The nomogram provides an easy-to-use clinical diagnostic tool and may be useful for preoperatively identifying SCNECC patients who might benefit from early and preoperative diagnosis through the determination of the most appropriate treatment plan. As a result, the improved SCNECC diagnosis method may benefit from cisplatin combined with etoposide chemotherapy that is recommended for SCNECC patients, which may reduce SCNECC mortality due to a delayed or an incorrect diagnosis. ^21,22

In the present study, the ADC histogram parameters displayed good discriminability between SCNECC from non-SCNECC, as shown by the following four observations:

First, the ADC_mean showed a significant difference (p < 0.001) between the SCNECC group (ADC_{mean =} 0.799×10⁻³ mm²/s) and the non-SCNECC (ADC_{mean =} 0.912×10⁻³ mm²/s) group. Duan et al ¹² reported that the ADC_mean value of cervical NECC was significantly lower than that of SCC and AC; the cut-off value was 0.900 × 10⁻³ mm²/s, which is different from our results, possibly because the more accurate three-dimensional (3D) whole-lesion ROI analysis was used in our study, whereas a two-dimensional (2D) ROI lesion analysis was used in their study. Moriya et al ²³ noted that the arbitrariness of setting the ROI may increase bias and that whole-lesion ADC values were more representative of tumour heterogeneity, which supports our argument that the whole-lesion three-dimensional (3D) ROI setting in this study may have yielded a more accurate diagnosis.

Second, ADC values are inversely related to tumour cellularity. ^24,25 The ADC_Minimum and low-percentile ADCs may reflect the hypercellular components in the tumour. ²⁶ Small cell neuroendocrine carcinomas are characterized by high cellularity with the presence of small blue cells with very scant or null cytoplasm. ²⁷ Our study showed that the ADC_Minimum (p < 0.001) and low-percentile ADCs (ADC_5th: p < 0.001; ADC_20th: p = 0.015) were significantly lower in the SCNECC group than in the non-SCNECC group. ADC_Minimum (OR = 0.362, 95% CI: 0.118–0.980, p = 0.0565) could be an independent diagnostic factor for SCNECC.

Third, the two histogram parameters ADC_SD (standard deviation) and ADC_{Inhomogeneity} (ADC_SD divided by the mean) quantitatively reflect the homogeneity of cellularity distribution in tumours. In our study, both ADC_SD (p = 0.001) and ADC_{Inhomogeneity} (p < 0.001) showed good performance in identifying SCNECC. This was probably related to an increase in diversification of cellularity, which might have been caused by different subtypes of SCNECC. Matoba et al ²⁸ indicated that the degree of necrosis and/or microstructural change that precedes necrosis may be a factor that influences ADC values in small cell lung carcinoma. The change can be reflected in the parameters of ADC inhomogeneity, which can be observed on T ₂-weighted images ²⁹ in SCNECC.

Fourth, some studies have reported that ADC_Skewness and ADC_Kurtosis are useful for differentiating cervical cancers from normal cervical tissues ³⁰ ; however, Guan Y et al reported that ADC_skewness was limited in its ability to differentiate cervical adenosquamous carcinoma from squamous cell carcinoma. ³¹ Our research results were similar to Guan Y et al.’s study. In the present study, ADC_Skewness (0.86 vs 0.84, p = 0.833) and ADC_Kurtosis (1.22 vs 0.94, p = 0.492) were not significant in differentiating SCNECC from non-SCNECC. This may be because both SCNECC and non-SCNECC are malignant tumours that are relatively inhomogeneous.

In addition to the ADC histogram parameters, the MR morphological features revealed in this study were also good indicators for discriminating SCNECC from non-SCNECC, as demonstrated by the following three findings:

First, tumour size is an important prognostic factor in cervical cancer, but there was no significant difference in tumour size between the SCNECC and non-SCNECC groups (p = 0.274) in this study. Duan et al ¹² also reported no significant difference (p = 0.763) in tumour size as reflected by MR results between NECC, SCC and AC. However, the tumour size analysis in this study showed a significant difference (p = 0.002) for tumour size >5 cm between the SCNECC and non-SCNECC groups, suggesting that a tumour size greater than 5 cm was a risk factor for SCNECC. Nonetheless, more large-scale research will be needed in the future.

Second, the SCNECC group showed a larger LN short axis than the non-SCNECC group (p = 0.001). This is in line with the clinical observation that SCNECC can easily invade the lymph nodes. This finding supports the hypothesis that it may be easier for SCNECC tumour cells to acquire an immune escape mechanism to avoid the host immune system. ³²

Third, primary cervical cancers with small tumour sizes are accompanied by larger metastatic lymph nodes, which is another imaging characteristic observed in this study. The LN-tumour ratio showed a significant difference (p = 0.001) between SCNECC and non-SCNECC, and therefore could be considered a risk factor for SCNECC.

Finally, clinical features also played an important role in distinguishing SCNECC and non-SCNECC. Takayanagi et al ³³ reported that 76% of patients with NECC were positive for HPV18. Alejo et al also reported that 41% of NECC patients had HPV18, while only 10% of patients with other histological types of cervical cancer had HPV18. ³⁴ The results of this study are similar in that there was a significant difference (p < 0.001) between the SCNECC (60%) and non-SCNECC groups (13%) in terms of HPV18 positivity. Moreover, HPV18 positivity was significantly more common among the lymphatic permeation predictors of relapse. ³⁵ This might explain why SCNECC has a high rate of lymph node metastases.

Due to the rarity of SCNECC, the patient pool was not as large as we had desired. Nevertheless, it was sufficient to allow meaningful statistical analyses for our model-building process. The power of the combined model and its corresponding nomogram in SCNECC diagnosis may be further improved if resources allow us to incorporate data from more SCNECC patients into the model-building process.

This study has some limitations. As a retrospective multicentre study, selection conditions were needed to ensure a relatively homogeneous patient sample. However, some of the selection conditions, such as the need for a 1.5 T magnetic resonance field, reduced the number of cases in this study. Partially due to the rarity of SCNECC, the sample size of the present study was relatively small, making it difficult to conduct subgroup analysis. External validation would be very beneficial for further generalization of the clinical use of the nomogram.

Conclusions

In conclusion, the nomogram built by whole lesion ADC histogram parameters combined with clinical and morphological features may serve as a non-invasive, reliable tool for individualized preoperative SCNECC prediction in clinical practice.