An MRI radiomics-based model for the prediction of invasion of the lymphovascular space in patients with cervical cancer

doi:10.21203/rs.3.rs-3604711/v1

Download PDF

Research Article

An MRI radiomics-based model for the prediction of invasion of the lymphovascular space in patients with cervical cancer

https://doi.org/10.21203/rs.3.rs-3604711/v1

This work is licensed under a CC BY 4.0 License

Version 1

posted

You are reading this latest preprint version

Background

Cervical cancer (CC) remains the second deadliest cancer-associated cause of mortality among women, and the ability to adequately predict the presence or absence of lymphovascular space invasion (LVSI) is vital to ensuring optimal patient outcomes. The objective of this study was to establish and verify an MRI radiomics-based model for the purpose of predicting the status of LVSI in patients with CC.

Methods

The present study conducted a retrospective analysis, wherein a total of 86 patients were included in the training cohort, and 38 patients were involved in the testing group, specifically focusing on patients with CC. The radiomics feature extraction process involved the utilization of ADC, T2WI-SPAIR, and T2WI sequences. Training group data were utilized for initial radionics-based model development, and model predictive performance was then validated based on data for patients enrolled in the experimental group.

Results

Radiomics scoring model construction was performed using 17 selected features. The study identified several risk variables associated with LVSI. These risk factors included elevated combined sequence-based radiomics scores (P < 0.001), more advanced FIGO staging (P = 0.03), cervical stromal invasion depth of a minimum of 1/2 (P = 0.02), and poorer tumor differentiation (P < 0.001). Radiomics scores based on combined sequences, ADC, T2WI-SPAIR, and T2WI exhibited AUCs of 0.931, 0.839, 0.815, 0.698, and 0.739 in the training cohort, respectively, with corresponding testing cohort values of 0.725, 0.692, 0.683, 0.833, and 0.854. The calibration curve analyses demonstrated an enhanced level of agreement between the actual and predicted LVSI status, indicating excellent consistency. Furthermore, the results of the decision curve study provided evidence for the clinical utility of this prediction model.

Conclusions

An MRI radiomics model was successfully developed and validated as a tool capable of predicting CC patient LVSI status, achieving high levels of overall diagnostic accuracy.

MRI

Radiomics

Lymphovascular space invasion

Cervical cancer

At the global level, cervical cancer (CC) remains the second deadliest form of cancer among women [1], with an estimated 530,000 diagnoses and 250,000 deaths annually [2]. Strikingly, an estimated 80% of CC-associated mortality impacts patients in developing nations [2]. Despite recent therapeutic advances, long-term CC patient outcomes remain unsatisfactory [3].

The presence of lymph node metastasis (LNM) is closely associated with CC patient prognosis [4–6], and evidence of lymphovascular space invasion (LVSI), in turn, is closely related to the LNM status of a given page. In patients with early-stage CC without LNM, positive LVSI status is indicative of the need for lymph node dissection [7]. Postoperative radiotherapy is also recommended in some LVSI-positive CC cases as per the Sedlis criteria [8]. The ability to preoperatively gauge the LVSI status of a given patient is thus vital to support therapeutic decision-making, yet only postoperative pathological examinations can currently confirm the presence of LVSI. Magnetic resonance imaging (MRI) approaches are widely used to diagnose CC as they can provide a high level of soft tissue resolution, but conventional MRI scans are poorly suited to the assessment of patient LVSI status [9–11].

Radiomics methods provide an alternate method for characterizing microscopic features that are invisible to the human eye by analyzing clinical images to obtain high-dimensional quantitative data [12]. MRI radiomics approaches have been deployed to predict tumor staging, LNM status, and therapeutic efficacy in patients with CC [4, 13–15]. No published MRI radiomics studies to date, however, have explored the preoperative prediction of CC patient LVSI status.

In order to address this knowledge gap, the present work was designed to develop and authenticate an MRI radiomics-based model capable of predicting LVSI status in cervical cancer patients.

Study design

This retrospective investigation was authorized by the Ethics Committee of Xuzhou Central Hospital, with the requirement for informed consent waived by the ethics committee. In this study, a total of 86 consecutive patients with CC who underwent an MRI examination between June 2016 and June 2021 were included in the training group. Additionally, a separate experimental group consisted of 38 consecutive CC patients who underwent MRI evaluation between July 2021 and October 2022.

The inclusion criteria were set as follows: (1) individuals having preoperative diagnoses confirmed by hysteroscopy, (2) had conventional MRI and DWI scans performed within 7 days before undergoing surgical treatment, and (3) underwent hysterectomy and pelvic lymph node dissection. Patients were excluded from these analyses if (1) their clinical data were not complete, (2) they exhibited any other comorbid malignancies, or (3) they had undergone preoperative chemotherapy or radiotherapy.

Data collected from all study subjects included baseline age, BMI, tumor differentiation, 2018 FIGO clinical staging, depth of invasion, and serum tumor biomarker levels (such as carcinoembryonic antigen [CEA], alpha-fetoprotein [AFP], carbohydrate antigen 199 [CA199], and squamous cell carcinoma antigen [SCC]).

MRI analyses

The MRI scanning procedure was conducted utilizing a 1.5T MRI apparatus manufactured by Philips, employing a body array coil known as Ingenia. All subjects were subjected to axial T2WI, fat suppression T2WI, axial diffusion-weighted imaging (DWI), sagittal T1WI, and sagittal T2WI sequence. The SPAIR sequences were utilized for fat suppression. For further details with respect to individual scanning protocols, see Supplemental Table 1.

Feature extraction

The axial ADC, axial T2WI-SPAIR, and Sagittal T2WI MRI results were analyzed with 3D Slicer (v 5.03) in which volumes of interest (VOIs) were defined by manually drawing the boundaries of target tumors (Fig. 1). The VOIs were restricted by two radiologists, each possessing 10 and 5 years of expertise, respectively. These radiologists were unaware of the pathological findings of the patients. Feature extraction was also performed using the 3D Slicer program, and the observer consistency was assessed based on both intra- and inter-class coefficient (ICC) values. Two radiologists performed Independent segmentation of MRI images from 20 randomly selected training cohort patients. In addition, one of these radiologists (Reader 1) repeated the segmentation of tumors from these same 20 patients 1-week after the initial segmentation. Those features for which an ICC ≥ 0.8 was measured were considered sufficiently repeatable and were retained for subsequent analyses. Reader 1 then segmented all remaining images.

Feature selection

To select features, a variance threshold approach was initially employed to identify those features exhibiting a variance > 0.75. The retained features were then subjected to the Selec-K-Best method, retaining those features with a P-value > 0.05. Finally, a LASSO regression model was used to select features related to LVSI status.

Radiomics model formulation and testing

The chosen features were employed to formulate a radiomics signature model, enabling the calculation of radiomics scores for each CC patient as follows:

Risk factors associated with the LVSI status of training cohort patients were selected via univariate and multivariate logistic regression approaches to enable the combined assessment of clinical features, serum biomarkers, and radiomics scores. Based on these results, a predictive nomogram was then developed, with subsequent validation being performed using data from the testing cohort.

Statistical analyses

Data were analyzed with SPSS 25.0 and R 4.1.2, comparing categorical data with χ² tests or Fisher's exact test. Normally (non-normally) distributed continuous data were compared with Student's t-tests (Mann-Whitney U tests). Risk factors related to LVSI status were selected through a series of logistic regression analyses. The DeLong test was employed to compare the area under the curve (AUC) values for generated receiver operating characteristic (ROC) curves.

Experimental group

The experimental group consisted of 86 patients diagnosed with cervical cancer, with 62 testing negative and 24 testing positive for LVSI. The data from these patients are summarized in Table 1. These groups exhibited comparable age, BMI, FIGO staging, pathological CC type, and serum tumor marker concentration distributions. However, the cervical stromal invasion depth (P = 0.007) and differentiation level (P = 0.001) did differ significantly between LVSI-positive and LVSI-negative patients.

Table 1

Baseline data of the patients in the training and test groups.
	Training (n = 86)			Test (n = 38)			P value*
	LVSI (-)	LVSI (+)	P value	LVSI (-)	LVSI (+)	P value	P value*
Patients’ number	62	24	-	30	8	-	-
Age (y)	57.5 (51.2, 64.8)	52 (47.2, 57.5)	0.068	53.4 ± 12.7	53.2 ± 6.8	0.964	0.214
BMI	22.6 ± 2.6	22.2 ± 2.7	0.541	23 ± 2.9	21.8 ± 1.5	0.109	0.612
Differentiation			0.001			0.137	0.508
Poor	6 (9.7%)	13 (54.2%)		3 (10%)	2 (25%)
Moderate	26 (41.9%)	10 (41.7%)		13 (43.3%)	5 (62.5%)
Well	30 (48.4%)	1 (4.2%)		14 (46.7%)	1 (12.5%)
Cervical stromal invasion depth			0.007			0.031	0.005
< 1/2	32 (51.6%)	4 (16.7%)		24 (80%)	3 (37.5%)
≥ 1/2	30 (48.4%)	20 (83.3%)		6 (20%)	5 (62.5%)
SCC			1.000			1.000	1.000
< 2.5 µg/L	22 (35.5%)	8 (33.3%)		10 (33.3%)	3 (37.5%)
≥ 2.5 µg/L	40 (64.5%)	16 (66.7%)		20 (66.7%)	5 (62.5%)
Ca199			0.496			1.000	1.000
< 37 U/L	52 (83.9%)	22 (91.7%)		25 (83.3%)	7 (87.5%)
≥ 37 U/L	10 (16.1%)	2 (8.3%)		5 (16.7%)	1 (12.5%)
AFP			0.057			0.309	1.000
< 7 µg/L	48 (77.4%)	23 (95.8%)		24 (80%)	8 (100%)
≥ 7 µg/L	14 (22.6%)	1 (4.2%)		6 (20%)	0 (0%)
CEA			0.203			0.587	0.573
< 5 µg/L	51 (82.3%)	16 (66.7%)		26 (86.7%)	6 (75%)
≥ 5 µg/L	11 (17.7%)	8 (33.3%)		4 (13.3%)	2 (25%)
FIGO stage			0.085			0.244	1.000
I	34 (54.8%)	8 (33.3%)		15 (50%)	3 (37.5%)
II	27 (43.5%)	14 (58.3%)		15 (50%)	4 (50%)
III	1 (1.6%)	2 (8.3%)		0 (0%)	1 (12.5%)
Pathological types			0.965			0.159	0.891
Adenocarcinoma	16 (25.8%)	7 (29.2%)		9 (30%)	0 (0%)
Others	46 (74.2%)	17 (70.8%)		21 (70%)	8 (100%)
AFP: α-fetoprotein; BMI: body mass index; CEA: carcinoembryonicantigen; FIGO: International Federation of Gynecology and Obstetrics; LVSI: lymphovascular space invasion; SCC: squamouse cell carcinoma antigen.
*: P values between training and test groups.

Feature selection

For each sequence (T2WI, T2WI-SPAIR, ADC), the extraction of 851 total radiomics features was performed. Results derived from the combination of these three sequences were then used for radiomics score establishment. For further details regarding the process of feature selection, see Supplementary Table 2. Ultimately, this strategy led to the selection of 17 features that were utilized for radiomics score calculations (Supplementary Fig. 1). THe corresponding coefficient values for these features and the combined sequence-based mean square errors are shown in Supplementary Fig. 2.

Predictive model development

In univariate analyses, factors associated with LVSI included elevated combined sequence-based radiomics scores (P < 0.001), advanced FIGO staging (P = 0.04), cervical stromal invasion depth ≥ 1/2 (P = 0.01), and a poorer differentiation (P < 0.001). Multivariate analyses confirmed that all of these factors were associated with LVSI incidence (Table 2).

Table 2

Risk factors of the LVSI.
	Univariate analysis			Multivariate analysis
	OR	95% CI	P value	OR	95% CI	P value
Age	0.97	0.93–1.01	0.2
BMI	0.94	0.79–1.13	0.53
Differentiation	0.14	0.06–0.35	< 0.001	0.14	0.05–0.37	< 0.001
Cervical stromal invasion depth	5.33	1.63–17.42	0.01	6.29	1.37–28.85	0.02
SCC	1.1	0.41–2.98	0.85
Ca199	0.47	0.1–2.34	0.36
AFP	0.15	0.02–1.2	0.07
CEA	2.32	0.8–6.76	0.12
FIGO stage	2.44	1.03–5.8	0.04	3.61	1.16–11.2	0.03
Pathological types	0.84	0.3–2.41	0.75
Radiomics score of combined sequences	419.79	20.07-8781.68	< 0.001	383.81	10.59-13913.67	< 0.001
AFP: α-fetoprotein; BMI: body mass index; CEA: carcinoembryonicantigen; FIGO: International Federation of Gynecology and Obstetrics; LVSI: lymphovascular space invasion; SCC: squamouse cell carcinoma antigen.

Based on these findings, a predictive model and nomogram were next established (Fig. 2). Nomogram scores were calculated with the following formula: score = -1.8306–2.1367×differentiation level (0: poor; 1: moderate; 2: well) + 2.1529 × cervical stromal invasion depth (0: < 1/2; 1: ≥ 1/2) + 1.482 × FIGO stage (0: I; 1: II; 2: III) + 7.7203 × combined sequence radiomics score.

The predictive model exhibited an AUC of 0.931, with respective sensitivity and specificity values of 83.9% when analyzing CC patients included in the training cohort (Fig. 3A, Table 3). This AUC value was significantly higher than corresponding values based solely on radiomics scores based upon the T2WI (0.739), T2WI-SPAIR (0.698), ADC (0.815), and combined sequence (0.839) features (P = 0.001, 0.001, 0.017, and 0.07).

Table 3

Diagnostic performance of each parameter.
	Training group			Test group
	Sensitivity	Specificity	AUC	Sensitivity	Specificity	AUC
Radiomics score of T2WI	70.8%	67.7%	0.739	75%	43.3%	0.725
Radiomics score of T2WI-SPAIR	79.2%	56.5%	0.698	87.5%	36.7%	0.692
Radiomics score of ADC	75%	77.4%	0.815	50%	63.3%	0.683
Radiomics score of combined sequences	87.5%	79%	0.839	87.5%	56.7%	0.833
Predictive model in this study	83.3%	83.9%	0.931	62.5%	76.7%	0.854
ADC: apparent diffusion coefficient; AUC: area under the curve.

Testing cohort validation

The 38 CC patients enrolled in the testing cohort included 30 and 8 patients with LVSI-negative and LVSI-positive disease, respectively. Of the analyzed baseline parameters, only cervical stromal invasion depth differed significantly between these two patient groups (Table 1). When the data from these patients were introduced into the developed predictive model, the respective AUC, sensitivity, and specificity values were 0.854, 62.5%, and 76.7% (Fig. 3B, Table 3). While the predictive model-derived AUC value was higher than corresponding values for T2WI (0.725), T2WI-SPAIR (0.692), ADC (0.683), and combined sequence (0.833) based radiomics scores, these differences were not statistically significant (P = 0.341, 0.118, 0.140, and 0.827, respectively).

Evaluation of Predictive Model Clinical Utility

In a calibration curve analysis, the model-predicted LVSI status of CC patients was very consistent with their actual LVSI status in the testing and training cohorts (Fig. 4). When a decision curve analysis was performed for this predictive model, the model yielded significant net benefits in the experimental and control groups, with corresponding risk thresholds of 0-0.97 and 0-0.38 (Fig. 5).

In this study, a model based on MRI radiomics was effectively created and validated as a predictive tool for the occurrence of lymphovascular space invasion (LVSI) in cervical cancer patients. A nomogram was ultimately established by combining MRI-based radiomics scores and relevant clinical features, and the resultant model exhibited high AUC values in training and testing cohorts consisting of individuals with CC (0.931 and 0.854, respectively). The data presented in this study indicate that the model under investigation has the potential to be successfully utilized in guiding healthcare choices and therapeutic planning for patients with CC.

In patients with early-stage CC, the ability to preoperatively gauge LVSI status is vital as the presence or absence of LVSI determines the primary treatment approach for individuals with clinical-stage IA disease [16]. Specifically, stage IA CC patients negative for LVSI can undergo curative cervical conization while preserving fertility [16]. Preoperative analyses of LVSI status can also minimize the risk of unnecessary lymphatic debridement in patients negative for LNM, thereby reducing postoperative complication-related risk [17].

Conventional MRI scanning provides an effective means of assessing tumor anatomy, morphology, and molecular movement [18]. Such scans, however, are poorly suited to LVSI detection as they are limited in their ability to evaluate microscopic tumor pathological findings [17], and no LVSI-related direct MRI feature has been identified.

As a means of analyzing tumors in greater detail, conventional images can be analyzed to extract radiomics features that correspond to a wide array of biological characteristics [17]. Li et al. [19] previously developed an MRI radiomics-based model aimed a predicting CC patient LSVI status, but the AUC values for their training and testing cohorts were just 0.754 and 0.727. These AUC values were likely lower than those in the present analysis because their prior study only utilized T1WI sequences for feature extraction [19].

Here, multimodel MRI sequences were employed for radiomics feature extraction. The predictive model that was generated produced an AUC value of 0.931 while evaluating patients in the experimental group. Additionally, the sensitivity and specificity values were determined to be 83.3% and 83.9%, respectively. The present study demonstrates the significance of a combined predictive radiomics scoring model, which surpasses the performance of radiomics scores produced from separate MRI sequences, as indicated by a greater AUC value. The use of multimodal MRI radiomics features can thus offer a higher degree of insight regarding the biological status of a given tumor relative to any individual MRI sequence. The final predictive model also incorporated differentiation level, cervical stromal invasion depth, and FIGO stage as clinical features closely related to the degree of CC malignancy. By incorporating these variables, the model demonstrated enhanced proficiency in anticipating the LVSI status of CC patients. In the experimental group, it was shown that the AUC for the predictive model did not exhibit a statistically significant increase compared to the AUCs obtained from radiomics scores derived from the concatenated sequences, namely ADC, T2WI, SPAIR, or T2WI. This may be attributable to the smaller sample size in this testing cohort.

This study is subject to some limitations. For one, the use of a retrospective patient cohort entails a high risk of bias, emphasizing a need for prospective validation. In addition, the relatively small number of enrolled patients inevitably limits the statistical power of these analyses. As tumor boundaries were manually defined, these results are also inevitably subject to some degree of measurement error. Lastly, the developed predictive model was based on a logistic regression analysis, and it may not be the most optimal predictive model given that machine learning models such as k-nearest neighbors, support vector machine, XG boost, random forest, or light GBM approaches were not employed.

In conclusion, a radiomics-based model utilizing MRI was effectively constructed and employed to forecast the state of LBSI in CC patients, demonstrating a commendable level of diagnostic accuracy across all participants.

ADC: apparent diffusion coefficient;

AFP: alpha-fetoprotein;

AUC: area under curve;

CA199: carbohydrate antigen 199;

CC: cervical cancer;

CEA: carcinoembryonic antigen;

DWI: diffused weighted imaging;

FIGO: International Federation of Gynecology and Obstetrics;

LASSO: least absolute shrinkage and selection operator;

LNM: lymph node metastasis;

LVSI: lymphovascular space invasion;

MRI: magnetic resonance imaging;

ROC: receiver operator characteristic;

SCC: squamous cell carcinoma antigen;

VOI: volume of interest.

Ethics approval and consent to participate: This study was approved by Ethics Committee of Xuzhou Central Hospital. The need for written informed consent was waived by the Ethics Committee of Xuzhou Central Hospital. All methods were carried out in accordance with Declaration of Helsinki.

Consent for publication: Not applicable.

Availability of data and materials: The data that support the findings of this study are available from the corresponding author upon reasonable request.

Competing interests: None.

Funding: None.

Authors’ contributions: SDL designed this work. NNM, and TW collected the clinical data. NNM performed the statistical analyses. NNM wrote this article. Final manuscript was approved by all authors.

Acknowledgments: None

Siegel RL, Miller KD, Jemal A. Cancer statistics, 2019. CA Cancer J Clin 2019;69:7-34.
Bray F, Ferlay J, Soerjomataram I, et al. Global cancer statistics 2018: GLOBOCAN estimates of incidence and mortality worldwide for 36 cancers in 185 countries. CA Cancer J Clin 2018;68:394-424.
Togami S, Kamio M, Kobayashi H, et al. Is It Possible to Perform Less Radical Surgery for Invasive Uterine Cervical Cancer? Gynecol Obstet Invest 2016;81:251-255.
Xiao M, Ma F, Li Y, et al. Multiparametric MRI-Based Radiomics Nomogram for Predicting Lymph Node Metastasis in Early-Stage Cervical Cancer. J Magn Reson Imaging 2020;52:885-896.
Kan Y, Dong D, Zhang Y, et al. Radiomic signature as a predictive factor for lymph node metastasis in early-stage cervical cancer. J Magn Reson Imaging 2019;49:304-310.
Deng, YR., Chen, XJ., Xu, CQ. et al. A preoperative nomogram predicting risk of lymph node metastasis for early-stage cervical cancer. BMC Women's Health 2023;23:568.
Wu Y, Wang S, Chen Y, et al. A Multicenter Study on Preoperative Assessment of Lymphovascular Space Invasion in Early-Stage Cervical Cancer Based on Multimodal MR Radiomics. J Magn Reson Imaging 2023;58:1638-1648.
Haesen J, Salihi R, Van Gorp T, et al. Radical hysterectomy without adjuvant radiotherapy in patients with cervix carcinoma FIGO 2009 IB1, with or without positive Sedlis criteria. Gynecol Oncol 2021;162:539-545.
Merz J, Bossart M, Bamberg F, et al. Revised FIGO Staging for Cervical Cancer - A New Role for MRI. Rofo 2020;192:937-944.
Balcacer P, Shergill A, Litkouhi B. MRI of cervical cancer with a surgical perspective: staging, prognostic implications and pitfalls. Abdom Radiol (NY) 2019;44:2557-2571.
Re GL, Cucinella G, Zaccaria G, et al. Role of MRI in the Assessment of Cervical Cancer. Semin Ultrasound CT MR 2023;44:228-237.
Mayerhoefer ME, Materka A, Langs G, et al. Introduction to Radiomics. J Nucl Med 2020;61:488-495.
Zhao X, Wang X, Zhang B, et al. Classifying early stages of cervical cancer with MRI-based radiomics. Magn Reson Imaging 2022;89:70-76.
Shi J, Dong Y, Jiang W, et al. MRI-based peritumoral radiomics analysis for preoperative prediction of lymph node metastasis in early-stage cervical cancer: A multi-center study. Magn Reson Imaging 2022;88:1-888.
Zhang Y, Zhang K, Jia H, et al. IVIM-DWI and MRI-based radiomics in cervical cancer: Prediction of concurrent chemoradiotherapy sensitivity in combination with clinical prognostic factors. Magn Reson Imaging 2022;91:37-44.
Bhatla N, Aoki D, Sharma DN, et al. Cancer of the cervix uteri. Int J Gynaecol Obstet 2018;143:22-36.
Luo Y, Mei D, Gong J, et al. Multiparametric MRI-Based Radiomics Nomogram for Predicting Lymphovascular Space Invasion in Endometrial Carcinoma. J Magn Reson Imaging 2020;52:1257-1262.
Zhang A, Song J, Ma Z, et al. Application of apparent diffusion coefficient values derived from diffusion-weighted imaging for assessing different sized metastatic lymph nodes in cervical cancers. Acta Radiol 2020;61:848-855.

No competing interests reported.

Download PDF

Version 1

posted

You are reading this latest preprint version

An MRI radiomics-based model for the prediction of invasion of the lymphovascular space in patients with cervical cancer

Status:

Version 1

Abstract

Background

Methods

Results

Conclusions

Figures

Introduction

Study methodology

Study design

MRI analyses

Feature extraction

Feature selection

Radiomics model formulation and testing

Statistical analyses

Results

Experimental group

Feature selection

Predictive model development

Testing cohort validation

Evaluation of Predictive Model Clinical Utility

Discussion

Conclusions

Abbreviations

Declarations

References

Additional Declarations

Supplementary Files

Status:

Version 1