A Hybrid Model-Based Clinicopathological Features and Radiomics Based on Conventional MRI for Predicting Lymph Node Metastasis and DFS in Cervical Cancer

Abstract

This study aimed to improve the accuracy of the diagnosis of lymph node metastasis (LNM) and prediction of patient prognosis in cervical cancer patients using a hybrid model based on MRI and clinical aspects. We retrospectively analyzed routine MR data from 485 patients with pathologically confirmed cervical cancer from January 2014 to June 2021. The data were divided into a training cohort (N = 261), internal cohort (N = 113), and external validation cohort (n = 111). A total of 2194 features were extracted from each ROI from T2WI and CE-T1WI. The clinical model (M1) was built with clinicopathological features including squamous cell carcinoma antigen, MRI-reported LNM, maximal tumor diameter (MTD). The radiomics model (M2) was built with four radiomics features. The hybrid model (M3) was constructed with squamous cell carcinoma antigen, MRI-reported LNM, MTD which consists of M1 and four radiomics features which consist of M2. GBDT algorithms were used to create the scores of M1 (clinical-score, C-score), M2 (radiomic score, R-score), and M3 (hybrid-score, H-score). M3 showed good performance in the training cohort (AUCs, M3 vs. M1 vs. M2, 0.917 vs. 0.830 vs. 0.788), internal validation cohorts (AUCs, M3 vs. M1 vs. M2, 0.872 vs. 0.750 vs. 0.739), and external validation cohort (AUCs, M3 vs. M1 vs. M2, 0.907 vs. 0.811 vs. 0.785). In addition, higher scores were significantly associated with worse disease-free survival (DFS) in the training cohort and the internal validation cohort (C-score, P = 0.001; R-score, P = 0.002; H-score, P = 0.006). Radiomics models can accurately predict LNM status in patients with cervical cancer. The hybrid model, which incorporates clinical and radiomics features, is a novel way to enhance diagnostic performance and predict the prognosis of cervical cancer.

Supplementary Information

The online version contains supplementary material available at 10.1007/s10278-024-01371-9.

Introduction

Cervical cancer (CC) is one of the most prevalent malignancies in women worldwide [1]. Lymph node metastasis (LNM) is an independent prognostic factor for CC, so the International Federation of Gynecology and Obstetrics (FIGO) released a new staging in 2018, with lymph node-positive patients classified as stage IIIC [2]. National Comprehensive Cancer Network (NCCN) guidelines recommend that radiotherapy and chemotherapy are the preferred treatment for CC patients diagnosed with LNM. However, there are still some patients who undergo surgical treatment and discovered LNM after surgery, requiring adjuvant radiotherapy and chemotherapy, which are associated with increased adverse effects and more complications [3]. Therefore, it is crucial to determine the status of lymph node by noninvasive investigations for selecting the most appropriate treatment option and avoid unnecessary surgery.

There are several methods to improve the accuracy of preoperative lymph node status assessment, including magnetic resonance imaging (MRI), positron emission tomography MRI (PET-MRI), and sentinel lymph node (SLN) biopsy, in which SLN provided the best diagnostic accuracy [4, 5]. However, the results of SLN biopsy are unstable due to a variety of external factors such as the experience of the clinician, rare lymphatic drainage patterns, etc. [6]. Although the diagnostic performance of PET-MRI is better than that of MRI, its low prevalence and high cost limit daily use [7]. MRI has been the preferred imaging method with for preoperative detection of LNM in CC patients [8]. However, false negative or positive results still occur in the detection of small metastatic lymph nodes as well as enlarged lymph nodes due to inflammatory hyperplasia [9]. Therefore, it is crucial to search for new biomarkers for the diagnosis of LNM in patients with CC.

Radiomics can analyze multiple quantitative imaging features by using different imaging modalities such as computed tomography(CT), PET, or MRI to create algorithmic models which correlate with clinical outcomes like diagnosis and prognosis [10, 11]. Radiomic features can quantify tumor intensity, shape and heterogeneity, which have been applied to diagnosis, treatment response, and prognosis in CC [12, 13]. Several studies have reported that radiomic features based on MRI and CT could be used to predict LNM in patients with CC [14–16]. Yan et al. explored the predictive value of an apparent diffusion coefficient (ADC)-based radiomics model for pelvic LNM in 153 patients with cervical squamous carcinoma, and the results showed that the radiomics model had good predictive performance in the training cohort and the validation cohort [17]. Lina et al. constructed column-line plots based on features extracted from T2 weighted imaging (T2WI), ADC, and contrast-enhanced T1 weighted imaging (CE-T1WI), and their hybrid model predicted LNM efficacy better than the model with separate sequences in training cohort and validation cohort [14]. Wu et al. found that intratumoral and peritumoral radiomics models from T2WI and ADC maps could predict LNM equally well, with AUCs greater than 0.8 in both training and validation cohorts [18]. Although there are some studies on the prediction of LNM, most of the studies use MR functional sequences such as diffusion-weighted imaging (DWI), ADC map, and intravoxel incoherent motion (IVIM), which are not as popular as conventional MRI. Since LNM is an independent factor influencing survival, preliminary results have been achieved that radiomics scores for LNM prediction can be used to predict patient prognosis in some studies [19, 20].

Hence, the study sought to validate radiomic signatures from conventional MRI for preoperative identification of LNM and assessment of individual disease-free survival (DFS) in patients with CC.

Materials and Methods

Patients

This retrospective study was approved by the Institutional Review Boards of the research center, and the requirement of informed consent was waived. Patients with CC from January 2014 to June 2021 were collected. The inclusion criteria were as follows: ① patients underwent radical hysterectomy and pelvic lymphadenectomy and ② patients with preoperative enhanced MRI. The exclusion criteria were as follows: ① patients receiving preoperative treatment (neoadjuvant chemotherapy, radiotherapy, or cervical conization); ② lesions less than 1 cm in MR images; and ③ other rare types of CC such as small cell carcinoma, clear cell carcinoma, and sarcoma.

Finally, 485 patients were included in this study (mean age, 51.5 ± 9.5 years; age range, 27–74 years). A flowchart of the study population is shown in Fig. 1. Three hundred seventy-four cases from Centre 1 were randomly divided into a training cohort (n = 261) and an internal validation cohort (n = 113) at a ratio of 7:3, and patients from Centre 2 as an external validation cohort (n = 111). The prevalence rates of LNM were 23.8%, 23.9%, and 23.4%, respectively, in three cohorts.

Fig. 1.

The patients’ recruitment pathway

The clinical characteristics included age, squamous cell carcinoma antigen (SCC-Ag), maximal tumor diameter (MTD), and MRI-reported LNM. The short diameter of the largest LN and pathology type of tumor were obtained by reviewing the medical records. All MRI scans were reviewed by two radiologists with 3 years and 18 years of experience. Based on commonly used criteria in daily clinical practice, patients with the short diameter of largest LN larger than 8 mm, round or irregularly shaped, rough edge of the lymph node or necrosis signal in the center were regarded as positive clinical LNM status [21]. The short diameter of the largest LN of each patient was recorded. MTD was determined by the longest diameter measured on sagittal or axial T2WI which were reviewed by a radiologist with 3 years of experience and then validated by a senior expert radiologist with 18 years of experience in imaging diagnosis of gynecology. Any disagreement was resolved by consultation.

MRI Acquisition

MRI examinations were performed on a 3.0 T or a 1.5 T MR scanner. The scanning sequence details are shown in Table 1.

Table 1.

Scan device parameter information

Hospital	Scanner	Sequence	Level	TR(ms)/TE(ms)	Layer thickness (mm)	Spacing (mm)	Matrix (mm)
Medical center one	3.0-T Verio Siemens	TSE T1WI	Axial	514/11	5	2	512 × 640
		TSE FS-T2WI	Axial	3000/88	5	2	294 × 448
		TSE T2WI	Sagittal	3500/116	5	1	360 × 448
		TSE CE-T1WI	Axial	677/11	5	2	432 × 640
			Sagittal	590/11	5	1.2	544 × 640
			Coronal	590/11	5	1.2	544 × 640
	1.5 T General Electric	TSE T1WI	Axial	440/8.47	7	1	512 × 512
		TSE FS-T2WI	Axial	4440/138.2	7	1	512 × 512
		TSE T2WI	Sagittal	3600/130.57	5	1	512 × 512
		GRE CE-T1WI	Axial	205/1.31	7	1	512 × 512
			Sagittal	200/1.62	5	1	512 × 512
			Coronal	185/1.59	5	1	512 × 512
Medical center two	3.0-T Trio Siemens	TSE T1WI	Axial	562/10	5	1	212 × 171
		TSE FS-T2WI	Axial	4326/102	5	1	288 × 254
		TSE T2WI	Sagittal	2725/98	5	1	308 × 308
		GRE CE-T1WI	Axial	3.4/1.32	3	1.5	200 × 171
			Sagittal	3.2/1.15	7	3.5	160 × 160
			Coronal	3.2/1.15	7	3.5	160 × 160

Follow-Up

Follow-up was conducted every three month until the 2nd year after treatment, twice per year in the 3rd, 4th, and 5th year, and once a year after that. Disease recurrence was confirmed by gynecological examination, tumor marker measurements, and imaging modalities such as CT, MRI, and PET-CT. In addition, DFS was chosen as the separate endpoints. DFS was defined as the time from the date of end of surgery to disease recurrence, death, or the last follow-up visit.

Image Segmentation

ITK-SNAP was used for manual 3D segmentation of axial FS-T2WI and sagittal CE-T1WI. Regions of interest (ROIs) were manually delineated by a fellow (with 3 years of experience), who was blinded to clinical outcomes. The ROIs were drawn along the margin of the tumor on each slice avoiding obvious necrotic regions. To evaluate the inter-observer reproducibility of ROIs, 50 patients were randomly selected for tumor segmentation by another radiologist with 5 years of experience in gynecological imaging. The sketching process of ROI is shown in Fig. 2.

Fig. 2.

A–D Patients without LNM, axial FS-T2-weighted (A, B) and sigittal contrast-enhanced T1-weighted (C, D) images; E–H patients with LNM. Axial FS-T2-weighted (E, F) and sigittal contrast-enhanced T1-weighted (G, H) images

Image Preprocessing

Before extracting the radiomics features, preprocessing of the MR images was performed, including normalization, resampling, discretization, and filtering of the images. Normalization was performed based on the gray values of each MR image. The parameter “bin Width” was set to 25 while making a histogram and discretizing the gray level of the images. The parameter “L OG Kernel size” was set to 1. The parameter “Resampling the voxel size” was set to 2. Afterward, the parameter “Based on the Baud sign” was set to be true. The parameter “Split object tag” was set to 1.

Feature Extraction

The Neusoft Imagingomics Platform was used to calculate the radiomics features. These radiomics features included Grey Level Correlation Matrix (GLDM), Grey Level Covariance Matrix (GLCM), shape-based, First-order statistical features (First order), Grey Level Tour Matrix (GLRLM), and Grey Level Region Size Matrix (GLSZM). In total, 2194 radiomics features were extracted from the axial FS-T2WI and sagittal CE-T1WI of each patient,respectively (Additional details of 2194 features can be found in the Annex 1).

Radiomics Feature Selection

A three-step feature selection used to reserve the significant features that are highly associated with LNM. First, the agreement between the two radiologists was assessed by the intraclass correlation coefficient (ICC). Features with an ICC higher than 0.75 were reserved. Finally, after analysis of L1-Logistic and least absolute shrinkage and selection operator (LASSO) regression, 4 features were selected.

Model Building

The clinical model (M1) was construct by Gradient Boosted Decision Tree (GBDT). After radiomics feature selection, the selected features were incorporated into GBDT, the radiomics model (M2) was developed. A hybrid model (M3) based on radiomics features consist of M2 and clinical features consist of M1 with the coefficients weighted by GBDT model. Three models were used to further analyze the survival status of patients according to the scores of M1 (clinical-score, C-score), M2 (radiomic score, R-score), and M3 (hybrid-score, H-score) for each patient. According to an optimal cutoff value of C-score, R-score, and H-score, the patients were classified into high- and low-score cohorts in the training cohort and internal validation cohort. DFS was chosen as the separate endpoints.

Statistical Methods

SPSS 26.0 was used for statistical analysis of the clinical and imaging indicators. Kolmogorov–Smirnov tests were used to determine whether the measurement data (age SCC-Ag, MTD) complied with a normal distribution and independent sample T inspection or rank-sum inspection were used to assess the metrological data. The chi square test was used to compare the count data (pathology type, MRI-reported LNM). The area under curve(AUC) of the receiver operating characteristic (ROC) was used to evaluate the performance of the M3 and compared with Delong’s test. Calibration curves were drawn to assess the consistency of M3 predicted and actual results using the “rms” package in R. Decision curve analysis (DCA) was performed to evaluate the net clinical benefits of the models using the “rmda” package in R. The cutoff value of the C-score, R-score, and H-score was determined by receiver operating characteristics curve analysis using the Youden index. Patients were then subdivided into low-risk and high-risk cohorts according to the cutoff value. Survival curves were generated using the Kaplan–Meier method. All statistical analyses were performed using SPSS 26.0 and R software version 3.6.3. P-value < 0.05 was considered statistically significant. The flow of this study is shown in Fig. 3.

Fig. 3.

Flowchart of this study

Results

Clinical Characteristics

Among the 485 patients, 115 patients had pathologically proven LNM, and 370 patients were pathologically confirmed to have no LNM. The clinicopathologic characteristics of patients in the training, internal, and external validation cohort were listed in Table 2. Significant differences were found in MRI-reported LNM, MTD on MRI, and SCC-Ag between the LNM-negative and LNM-positive cohorts. These clinical indicators were used to build M1 based on GBDT. The example of LNM-negative versus LNM-positive patients are shown in Figs. 4 and 5.

Table 2.

Comparison of clinicopathologic features among training, internal, and external validation cohort

Characteristic	Training cohort n = 261		P	Internal validation cohort n = 113		P	External validation cohort n = 111		P
	LNM( +)62(23.8%)	LNM( −)199(76.2%)		LNM( +)27(23.9%)	LNM( −)86(76.1%)		LNM( +)26 (23.4%)	LNM( −)85(76.6%)
Age	50.5 ± 9.46	52.80 ± 9.14	0.080	50.50 ± 9.47	46.50 ± 11.00	0.066	50.81 ± 8.29	54.67 ± 9.54	0.066
SCC-Ag	2.40 (1.38, 6.00)	1.60 (0.80, 3.50)	0.002*	2.20 (1.00, 7.20)	1.35 (0.70, 2.63)	0.050	1.00 (0.58, 2.45)	1.00 (0.60, 1.85)	0.883
Pathology type
Squamous	55 (88.70%)	177 (88.94%)	0.583	22 (81.50%)	72 (83.70%)	0.630	21 (80.80%)	76 (89.40%)	0.245
Adenocarcinoma	7 (11.30%)	19 (9.55%)		5 (18.50%)	12 (14.00%)		5 (19.20%)	9 (10.60%)
Adenosquamous carcinoma	0 (0.00%)	3 (1.51%)		0 (0.00%)	2 (2.30%)		0 (0.00%)	0 (0.00%)
MRI-reported LNM			< 0.001*			< 0.001*			< 0.001*
Yes	23 (37.10%)	187 (94.00%)		15 (55.60%)	1 (4.70%)		15 (57.70%)	12 (14.10%)
No	39 (62.90%)	12 (6.00%)		12 (44.40%)	85 (95.30%)		11 (42.30%)	73 (85.90%)
MTD	39.00 (32.00, 45.25)	32.00 (24.00, 38.00)	< 0.001*	37.00 (32.00, 45.00)	31.00 (25.00, 39.00)	0.012*	34.00 (25.00, 43.25)	26.00 (20.00, 34.50)	0.010*
Short diameter of the largest LN	3.00 (0.00, 11.00)	4.00 (0.00, 6.00)	0.846	8.00 (0.00, 9.00)	4.00 (0.00, 6.00)	0.067	0.00 (0.00, 8.25)	4.00 (0.00, 6.00)	0.645

Fig. 4.

Patient 1, a 36-year-old female patient without LNM (A–D): A axial T2-weighted fat-suppression image, B axial contrast-enhanced T1-weighted sequence, C sagittal T2-weighted sequence, and D sagittal contrast-enhanced T1-weighted sequence; a maximal tumor diameter (MTD) of 2.4 cm, SCC-Ag0.8 ng/ml, C-score 0.095, R-score 0.179, and H-score 0.057, and no metastatic lymph nodes were found on MRI. No recurrence was observed by the end of follow-up

Fig. 5.

A 51-year-old female patient with LNM (A–D): A axial T2-weighted fat-suppression image, B axial contrast-enhanced T1-weighted sequence, C sagittal T2-weighted sequence, and D sagittal contrast-enhanced T1-weighted sequence; the short diameter of the largest LN 1.4 cm, poorly differentiated SCC (70 ng/ml), and a MTD of 3.2 cm, C-score 0.543, R-score 0.188, and H-score 0.748; relapse occurred at 36 months of follow-up

Feature Selection and Radiomics Model Building

In the radiomics features, 4 radiomics features were selected as the most representative radiomics features to construct M2 (Table 3), including 2 features from FS-T2WI (original_shape_LeastAxisLength and log-sigma-2–0-mm-3D_firstorder_80Percentile) and 2 features from CE-T1WI (wavelet-LLH_gldm_LargeDependenceHighGrayLevelEmphasis and log-sigma-2–0-mm-3D_gldm_LargeDependenceHighGrayLevelEmphasis). The radiomics scores of each patient were computed according to the filtered radiomic features. According to M2, the R-score of each patient in the training, internal, and external validation cohort was calculated and compared, and the results showed that the R-score of patients with LNM was higher than that of patients without LNM (Fig. 6).

Table 3.

Radiomic features

MRI sequence	Radiomic features
FS-T2WI	original_shape_LeastAxisLength
	log-sigma-2–0-mm-3D_firstorder_80Percentile
CE-T1WI	wavelet-LLH_gldm_LargeDependenceHighGrayLevelEmphasis
	log-sigma-2–0-mm-3D_gldm_LargeDependenceHighGrayLevelEmphasis

Fig. 6.

A–C R-score of three cohorts

Model Comparison and Validation

Figure 7 depicts the ROC curves of the different models. M3 showed the best performance in predicting LNM in the training, internal, and external cohorts, with AUCs of 0.917, 0.872, and 0.907, respectively. According to the Delong test, there was a significant difference between M1 and M3, M2 and M3 (P ≤ 0.001) (Table 4), and M3 was significantly better than the radiomics model (M2) or clinical model (M1). Subsequently, the calibration curve of the M3 demonstrated good calibration in the three cohorts (Fig. 8). The Hosmer–Lemeshow test yielded P values of 0.419, 0.674, and 0.638 in the training cohort, internal validation cohort, and external validation cohort. As shown in Fig. 9, the normalized confusion matrix further showed the M3 classification accuracy on the training cohort and internal validation cohort (1 represents LNM( +) and 0 represents LNM(-)). Decision curve analysis (DCA) of three models for prediction of LNM is presented in Fig. 10; the decision curve showed that if the threshold probability was > 4%, the use of the hybrid model to predict LNM added more benefits than either the radiomic signature or the clinical diagnosis.

Fig. 7.

A–C ROC curves of three cohorts

Table 4.

Quantitative indexes of three models on training and validation cohorts

	Training cohort			P	Internal validation cohort			P	External validation cohort			P
	AUC	Sen	Spe		AUC	Sen	Spe		AUC	Sen	Spe
M1	0.830	0.613	0.930		0.750	0.444	0.953		0.811	0.731	0.812
M2	0.788	0.823	0.618		0.739	0.889	0.523		0.785	0.808	0.647
M3	0.917	0.774	0.910		0.872	0.741	0.895		0.907	0.846	0.835
M1 vs M2				0.376				0.874				0.690
M1 vs M3				0.010*				0.079				0.035*
M2 vs M3				0.001*				0.031*				0.023*

Fig. 8.

A–C Calibration curves of three cohorts

Fig. 9.

Confusion matrix for the training cohort and internal validation cohort of M3

Fig. 10.

Decision curves of three models

Survival Prediction

As of 28 November 2023, the follow-up was completed for patients from the Medical Centre I. The overall recurrence rate was 9.09% (34/374), and the median disease-free survival of the patients was 52 months (range 3–79 months).

According to an optimal cutoff value of C-score, R-score, and H-score, the patients were classified into high- and low-score cohorts in the training cohort and internal validation cohort. Higher scores were significantly associated with worse DFS in the training cohort and the internal validation cohort (C-score P = 0.001, R-score P = 0.002, H-score P = 0.006) (Fig. 11). The 5-year DFS for high-risk and low-risk cohorts was 91.2% and 80.5% based on H-score, respectively. C-score, R-score, and H-score were independent predictors of disease-specific recurrence.

Fig. 11.

A–C Kaplan–Meier curve of C-score, R-score, and H-score

Discussion

According to the 2018 edition of the NCCN guidelines, LNM is an independent influencing factor for poor prognosis of cervical cancer, which significantly affects the selection of treatment options for early CC [22]. The prognosis of patients with CC is associated with the presence of LNM, which decreases the 5-year survival rate from 93.5 to 80.0% [23]. Therefore, accurate evaluation of LNM aids in designing treatment regimens for the patients with CC to improve their prognosis. In this study, a clinical-radiomics hybrid model was built for the preoperative prediction of LNM in CC patients, which consists three clinical risk factors (SCC-Ag, MTD, and MRI-reported LNM) and radiomics features extracted from the tumor regions in FS-T2WI and CE-T1WI achieved an AUC of 0.917. This finding is largely consistent with those reported in previous studies, indicating that the hybrid model was better than the single radiomics or clinical model. The M3 showed the best discriminative power in both the training (AUC: 0.917) and internal cohort (AUC: 0.872) and external cohort (AUC: 0.907) and indicated excellent discrimination ability in distinguishing between the patients who progress and those who do not progress at 5 years.

Clinical Characteristics

In this study, SCC-Ag generates a poor prognostic outcome, which is in line with a prior study that SCC-Ag increases the risk of LNM in CC [24]. SCC-Ag is a tumor marker for squamous cell carcinoma and an important biomarker for the disease severity and progression in patients with cervical squamous cancer. Although the patients with different pathological types of CC were collected in this study, cervical squamous cancer accounted for a relatively large number of the patients, and thus, the relationship between SCC-Ag and LNM was shown in the experiment. In addition, the MTD was confirmed to be positively correlated with the risk of LNM in this study. The larger maximum diameter, the more likely to develop LNM, which is similar to the results of other studies [21]. Probably the CC with the larger diameter has the more obvious invasion of surrounding tissues, and the more likely it is to undergo LNM. Currently, previous studies have shown that LNM status reported by MRI is an independent predictor of LNM in CC [25], which ties well with this study; the mostly accepted criterion on MRI in the evaluation of pelvic LNM is the size of the lymph node, with the short diameter of the LN larger than 8 mm indicating LNM [26]. However, the sensitivity of this criterion is extremely low with 30.3–54.0% in previous studies [27, 28], which is in consistent with the sensitivity of 46% in this study. Consistent with most studies, it suggests that a certain number of patients with LNM get the normal sized lymph nodes. There was no statistically significant difference in the short diameter of the largest LN included in this study between the LNM-negative and positive cohorts, which may be due to the fact that the diameters of some lymph nodes diagnosed as metastasis did not meet the diagnostic criteria of 8 mm, but these lymph nodes were diagnosed by radiologists as LNM because of morphological changes such as cystic necrosis or obvious irregular margins in the lymph nodes. Therefore, according to the current MRI diagnostic criteria, there are problems of high specificity and low sensitivity, and there is a high missed diagnosis rate for metastases without significant enlargement of lymph nodes.

Radiomics Features

The GBDT model cited in this study is one machine algorithms which has a greater advantage than the traditional logistic regression model in dealing with massive disordered clinical data because the modeling rules of the algorithm are not fixed and can be flexibly transformed, which is conducive to further promotion in clinical practice [29]. At present, GBDT has been applied to the prediction of prostate cancer, lung lesions, and other diseases, and the diagnostic efficiency is high, and the model prediction is stable [30, 31], whereas there is less prediction in the direction of cervical cancer.

The FS-T2WI and CE-T1WI are the most commonly used MR sequences for evaluating the staging of cervical cancer, T2WI can provide information about tumor morphology and tissue components, and CE-T1WI can reflect the microenvironment information of the tumor, which is helpful for the detection of lymph nodes [32]. Therefore, the FS-T2WI and CE-T1WI were selected as the sequences for radiomics feature extraction in this study. Four features of the 4388 radiomic features were identified to be predictive for LNM including 2 features extracted from FS-T2WI and 2 features derived from CE-T1WI, which are one original shape feature and one first-order feature based on LoG transformation, respectively, and others are Gray Level Dependence Matrix [33]. The one original shape feature from FS-T2WI is the least axis length indicating tumor size on T2WI, and one first-order feature based on LoG transformation from the FS-T2WI describes the voxel intensity distribution in the image area defined by the mask, which is related to the grayscale frequency distribution within the ROI and can directly reflect the morphology and pixel situation of the lesion. Two features from CE-T2WI are Gray Level Dependence Matrix, which reflects the grayscale relationship between the central pixel or voxel and the neighborhood and belongs to texture feature similar to first-order features [34]. Texture features can reflect the heterogeneity of tumors, and the higher heterogeneity within the tumor, the easier it is for tumors with larger grayscale differences to infiltrate the periphery. Both grayscale dependency matrices were extracted from the CE-T1WI, which might be due to the more significant differences between the internal component signals of the lesion on the enhanced sequence, and the more significant comparison between the lesion and its surrounding normal tissue, allowing for more accurate delineation of ROI.

Hybrid Model and Its Predictive Value

In this study, we developed a total of three LNM prediction models for cervical cancer patients, and M3 has the highest diagnostic efficiency. After adding radiomics features to the M1, the sensitivity of the training cohort increased from 0.613 to 0.774, and the AUC also increased from 0.830 to 0.917. This study found that radiomics properties extracted by MRI could enhance the diagnostic accuracy of LNM before surgery.

A study using the radiomics method to predict LNM in cervical cancer patients showed that the prediction performance of its radiomics signature was lower than ours, perhaps because the related serologic biomarkers of cervical cancer features such as SCC-Ag were not extracted [9]. Wu et al. [18] predicted LNM with an AUC of 0.747 based on radiomics features from multiparametric MRI including ADC and dynamic contrast-enhanced MRI. There are also some studies that have established multiparametric MRI-based radiomics nomogram for preoperative LNM status prediction [9, 14].

The use of conventional MRI sequences in this study showed diagnostic efficiency comparable to that of other research models (AUC 0.917), which may be due to the large number of patients in this study compared with other studies, the inclusion of clinical factors that more intuitively reflect tumor-related information, and the stability of the model has also been verified in multicenter patients, combining clinical factors and radiomics features to construct a hybrid model, which demonstrated high sensitivity (77.4%, 74.1%, and 84.6%) in the detection of LNM in training, internal, and external validation cohorts, with higher AUC and better calibration, consistent with previous studies [17, 35].

Implications of the LNM Prediction Model in the Prognosis of Patients with CC

LNM is significantly associated with the prognosis of CC patients [36, 37], and pelvic LNM is closely associated with the risk of death in patients with CC [35]. Pedro et al. [38] assessed the prognostic impact of clinical and pathological factors on patients with locally advanced CC and found that the patients with metastatic pelvic lymph nodes with a diameter of > 4.38 cm had a significant reduction in overall survival (HR 95.1;65% CI 116–0.003; P = 5.14). Yan et al. [39] studied 78 patients with IIIC1, and the patients with more than or equal to 2 LNM positions were found to have poorer DFS and OS at 5 years than 1 or 2 LNM positions. Luigi et al. [40] studied the prognostic significance of multiple pelvic LNMs in the patients with CC who underwent radical hysterectomy and bilateral pelvic lymphadenectomy, and they demonstrated that the patients with multiple LNMs (> 2) had a shorter DFS than the patients with 1 or 2 LNMs.

The study analyzed clinical indicators, radiomics features, and their combination as prognostic factors for LNM and prognosis and found that the C-score, R-score, and H-score could achieve the good prediction performance, and H-score with the best prediction performance, similar to the findings of other studies [12, 19]. The results mean the C-score, R-score, and H-score based on clinically relevant factors and radiomic features simultaneously reflect the invasiveness and survival of tumors. Radiomics can mine more prognostic information by observing the whole tumor scope and extracting high-dimensional features, which could be used as a surrogate biomarker to provide assistance for personalized treatment.

This study has several limitations. First, although the number of patients was larger as compared with previous studies, prospective datasets would be needed to optimize the performance of the model in the future. Second, due to the long data collection time for survival analysis, there is a lack of DWI images in early patients, and there is no correlation analysis, DWI sequences will be added in follow-up studies to improve diagnostic efficiency. Third, few patients in the external validation cohort had progression which did not allow to assess the prognostic effect of the model, and long-term follow-up of this cohort of patients is needed for the validation of the model’s effect. Finally, the exploration of the connection between radiomics features and genomic characteristics for LNM of cervical cancer was not considered in this study. Some preliminary studies showed that the combination of the omics mentioned above may improve the ability of disease diagnosis and treatment [41, 42].

Radiomics, since its inception in 2016, has seen an explosion of research, with a steady increase in the quantity and quality of research in this field; guidelines in field of radiomic were recently issued by the European Society of Radiology, aiming at standardizing the radiomic workflow, as this should be the first step to be implemented in order integrate radiomics data in clinical routine. We believe that enhancing the model’s user-friendliness will also improve its acceptability among medical professionals and patients, potentially facilitating its clinical adoption. In order to realize the full potential of radiomics in the evaluation of the efficacy of neoadjuvant cancer therapy, we need to design and conduct rigorous multicenter, large-sample clinical studies with prospective validation. Continued research is also needed to improve the interpretability of radiomics models. Overcoming these challenges would allow to accelerate the routine application of this innovative approach in the clinic to better serve patients.

Overall, the study presents that three models including clinical model, radiomics model, and hybrid model were closely related to prognosis while characterizing LNM in the patients with CC. A hybrid model based on clinical and radiomics features facilitates the preferable prediction of LNM in the patients with CC.

Supplementary Information

Below is the link to the electronic supplementary material.

Author Contribution

All authors contributed to the study conception and design. Material preparation, data collection, and analysis were performed by Mingke Tian. The first draft of the manuscript was written by Mingke Tian, and all authors commented on previous versions of the manuscript. All authors read and approved the final manuscript.

Funding

This work was supported by two grants from the Applied Basic Research Project of Liao Ning province (2022JH2/101300074) and Cancer Hospital of Dalian University of Technology (Liaoning Cancer Hospital & Institute) (LD2023034) (supported by “the Fundamental Research Funds for the Central Universities).

Data Availability

Data generated or analyzed during the study are available from the corresponding author by request.

Declarations

Ethics Approval

This is an observational study. The Research Ethics Committee has confirmed that no ethical approval is required.

Consent to Participate

Informed consent was obtained from all individual participants included in the study.

Consent for Publication

The authors affirm that human research participants provided informed consent for publication of the images in Figs. 2, 4, and 5.

Competing Interests

The authors declare no competing interests.