Application of Machine Learning for Prognostic Prediction in early-stage Cervical Cancer, Based on radiomics from ultrasound and MRI

doi:10.21203/rs.3.rs-3759028/v1

Download PDF

Research Article

Application of Machine Learning for Prognostic Prediction in early-stage Cervical Cancer, Based on radiomics from ultrasound and MRI

https://doi.org/10.21203/rs.3.rs-3759028/v1

This work is licensed under a CC BY 4.0 License

Version 1

posted

You are reading this latest preprint version

Purpose: We aimed to develop a model for an early-stage cervical cancer for disease free survival (DFS) prediction using machine learning methods based on the combination of clinicopathological and radiomic features which is extracted from magnetic resonance imaging (MRI) and ultrasonography (US).

Methods: This retrospectively study included 144 patients who were randomly divided into training and testing cohort at a ratio of 6:4.Radiomic features were extracted from MRI and US images, and in total, 1180 radiomic features and 9 clinicopathological factors were obtained. Six supervised machine learning classifiers were used to assess the prediction performance based on all variables. Next, we established models based on various combinations of clinicopathological characteristic and radiomic features to get the best prediction model using LightGBM. The model’s performance was evaluated by accuracy (ACC) and area under the curve (AUC). Furthermore, unsupervised clustering analysis was performed to identify CC patient subgroups related to DFS prognosis based on the all variables.

Results: LightGBM was superior to any other classifiers in CC DFS prediction. The model that combined clinicopathological factors with radiomic features from MRI and US showed the best performance, and the corresponding values were 0.92 of ACC and 0.86 of AUC. Unsupervised clustering analysis identified a strong tendency toward the formation of two distinct groups in DFS rate among CC patients.

Conclusion: MRI and US based radiomics has the potential of DFS prediction in early-stage CC with the LightGBM classifier, and the use of predictive algorithms may facilitate the personalized treatment options.

cervical cancer

radiomic

LightGBM

prognosis

ultrasound

MRI

Cervical cancer (CC) has been noted as the fourth most common cancer in women globally(Siegel et al., 2022), still being the primary cause of death in women in developing countries (Maguire et al., 2015). The National Comprehensive Cancer Network (NCCN) have been commonly used to determine the postoperative prognosis for early-stage of CC patients and postoperative concurrent chemoradiotherapy (CCRT) is suggested following for patients with adverse pathologic factors(Koh et al., 2019). Despite CCRT improves the survival rate of advanced cervical cancer, some patients were still treated with poor efficacy(Bhatla et al., 2019). Meanwhile, increased morbidity and complications has been indicated when surgery and postoperative CCRT are combined(Ayhan et al., 2004; Barter et al., 1989). The determinants of prognosis included not only pathological factors, but physical characteristics, such as age and body mass index (BMI), have also been proved to influence prognosis. (Huang et al., 2019; Papathemelis et al., 2019; Yu et al., 2021). Accurate prediction can not only help for the decision of CCRT regimen but also provide a guidance for change the drug type or adjust the radiation dose for patients who were insensitive to regular treatment (Haldorsen et al., 2019; Wang et al., 2020b). Therefore, accurate prediction is the great importance for physicians to determine a suitable treatment plan for reduction in incident of morbidity and improvement in quality of life after treatment.

Although CC have been diagnosed and followed up on the basis of combination of physical features and laboratory factors, noninvasive imaging examinations, including computed tomography (CT), magnetic resonance imaging (MRI) and ultrasonography (US), have also played important roles in recent years (Lu and Lu, 2022; Merz et al., 2020; Tian and Luo, 2022). With the advantages of higher spatial and contrast resolution of pelvic tissues and organs, MRI was considered as an important technique especially in assessing the depth of invasion and tumor volume, which were indicated as risk factors for recurrence and metastasis (Balcacer et al., 2019; Lu and Lu, 2022). Novel technologies, such as T2 weighted MRI (T2WI) and diffusion-weighted imaging (DWI), can provide additional tissue metabolic information and thus have the potential to be used in the staging of localized cervical cancer (Manganaro et al., 2021). US examination has long been universally recognized as a technique in diagnosis of CC due to its convenience, low-cost and radiation-free (Tian and Luo, 2022), but limited to its value of the depth of myometrial invasion (DMI) and lymph node status (LNM), and often relied on the skill of the US. But researches suggested that US can be useful as well as MRI in assessment of extension (Fischerova et al., 2008). Meanwhile, a new diagnostic method that integrates both US and MRI tests, was recently demonstrated for increasing the diagnostic accuracy of the single imaging method (Berger et al., 2018; Theodore et al., 2017) .

Traditional imaging features were obtained by human eyes on image grayscale, which has the limitation for more valuable microcosmic image feature identification. Radiomics, which converted digital medical images into quantitative, high-dimensional data, have been drawn much attention currently (Yu et al., 2020). It can effectively extract massive features from images and reveal innate characteristics of tumors for treatment response prediction. There were increasing published studies proving that radiomic features could be applied as significant factors for accurate diagnosis and prognosis prediction in tumors (Hou et al., 2020; Wang et al., 2020a; Yu et al., 2020), including cervical cancer(Hou et al., 2020; Wang et al., 2020a), which can facilitate clinical decision-making.

However, radiomic research may not reach the clinical useful level since they often based on conventional learning method, which was restrict by future bigger data management. Whereas, when coupled with machine Learning (ML) algorithm, which can analysis date-driven for mining implicit clinical values of radiomic features, radiomic served as imaging biomarkers to establish clinically diagnostic and predictive models, yielding some promising results (Castiglioni et al., 2019; Dias and Torkamani, 2019; Forghani et al., 2019), such as lymph node states prediction, survival or genetic mutation. Such a new algorithm was commonly admitted having two main advantages over conventional machine learning method: the first one is that ML performed much better when deal with larger samples; and the other one is that it could detect more complex and different types of biomarkers. However, no study as yet to date has established based on the combination of the radiomic from MRI and US, clinicopathological factors and ML algorithm in cervical cancer.

In this study, we aim to develop an early-stage cervical cancer–specific prediction method based on machine learning algorithms (Light Gradient Boosting decision machine [LightGBM]) using radiomics features from MRI and US and clinicopathological factors for detecting the predictive performance in patients with cervical cancer.

Patient data

This retrospective study was approved by the Ethics Committee of the First Affiliated Hospital of Wenzhou Medical University(number: 2021-zz-048), and the requirement of informed consent was waived. A total of 596 patients who underwent surgeries in our hospital with pathologically confirmed early-stage CC between September 2013 and August 2017 were firstly enrolled. Both ultrasound and MRI including T2WI and DWI were performed by experienced reference practitioners blinded to International Federation of Gynecology and Obstetrics (FIGO) in our hospital within two weeks before operation. The other inclusion criteria were that patients with FIGO stage less than or equal to IIA2 and patients underwent radical hysterectomy and pelvic lymphadenectomy. The criteria for exclusion were defined to include patients who underwent preoperative neoadjuvant chemotherapy or radiotherapy, suffered from concurrent malignancies at other sites, had incomplete clinicopathological parameters and could not proceed accurate texture analysis due to the tumor on the US images and MRI images was too small.

A total of 144 patients were finally included. Dataset was then split into training and testing cohorts with the ratio of 6:4, with repeated random sampling until no significant difference was seen between the two cohorts with respect to all features (p > 0.05 for all). This resulted in allocation of 87 patients to the training cohort and 57 patients to the testing cohort.

Retrieval of Clinicopathological information

For each patient, clinicopathological parameters were obtained from medical records, including age, BMI, FIGO stage, LNM, tumor size, pathological type, differentiation, DMI and lymph-vascular space invasion (LVSI) status. Regular follow-up was conducted every 3 months for the first two years, every 6 months during 3–5 years, and thereafter once a year. Disease free survival (DFS), which was defined as the period from operation to either the date of first recurrence or death, was the endpoint of our study. Recurrence is defined as biopsy-confirmed CC, gynecological examination or imaging exhibiting CC. The prognosis information of each patients was obtained by medical records system or telephone follow-up, and the last follow-up was conducted in October 2020.The complete list of clinicopathological parameters for each subject is summarized in Table 1.

Image acquisition, tumor segmentation and radiomics

Image acquisition and tumor segmentation

Pelvic MRI examination was performed with pelvic 3.0-T MRI scans, using a 16-channel phased-array encoding abdominal coil. The parameters were detailed described in the Supplementary Method 1. Regions-of-interest (ROIs) were manually segmented by a senior radiologist who had 10 years of experience in gynecological MRI imaging interpretation, and thereafter validated by the senior radiologist blinded to histologic information.

Patients underwent US by using ultrasonic equipment of HITACHI or PHILPS, and 5–14 MHz linear probes was adopted. To extract representative ultrasound image features, every image for each patient was selected and reviewed by a junior radiologist who has 8 years of experience in US and followed by a senior skilled radiologist who has more than 20 years of experience in gynecological tumors.

Radiomics feature extraction

Before feature extraction, the open-source ITK-SNAP software (http://www.itksnap. org) for three-dimensional manual segmentation in T2WI and DWI images and static images of the largest cross section in US were used. To reduce heterogeneity bias caused by different equipment and scanning parameters, the images were preprocessed by resampling the images to a 1 × 1 × 1 mm³ voxel size and normalizing the image gray scale to 0 to 255. The algorithm implemented in the artificial intelligent Kit version (3.0.0.R) was then used for radiomic feature extraction and based on which a total of 1180 radiomic features were obtained from ROI. The selected features were divided into five groups: (I) intensity histogram, (II) morphology, (III) gray-level co-occurrence matrix (GLCM), (IV) gray-level run length matrix (GLRLM), (V) gray-level size zone matrix (GLSZM).

In total, 1189 features were obtained in each patient, with 1180 radiomic features from MRI and US images and 9 clinicopathological factors.

Supervised Machine Learning Classifiers

After the image radiomics features extraction, all the clinicopathological and radiomics features were incorporated into a single model to form the training sample set. To improve the predictive performance, different machine learning algorithms were obtained, including Random Forest (RF), Support Vector Machine (SVM), LightGBM, Nave Bayesian (NB), K-Nearest Neighbor (KNN), and Logistic Regression (LRM). All classifiers were trained in the training cohort using repeated 10- fold cross-validation to minimize potential over-fitting, and then were tested on the testing cohort. The receiver operation characteristic (ROC) analysis, area under the curve (AUC) value and accuracy (ACC) were employed to validate the performance of these models. The complete workflow of the radiomic analysis implemented in this study is presented in Fig. 1.

LightGBM

LightGBM (Kobayashi and Yoshida, 2021) is a gradient boosting framework that uses tree-based learning algorithms. It has gained popularity for its speed, scalability, and high performance on large datasets.

LightGBM is based on the concept of boosting, where weak learners (i.e., decision trees) are sequentially added to a model to improve its performance. The boosting process involves adjusting the weights of the training examples to focus on the ones that are difficult to classify, and adding trees that can better handle these examples. LightGBM uses a gradient-based approach to tree learning, which can improve accuracy by handling non-linear relationships between variables and reducing overfitting. LightGBM can also provide feature importance rankings and visualization tools that help users understand the contribution of each variable to the model's predictions.

In recent years, k-fold cross-validation has been used in prediction, and therefore, ten-fold cross-validation is used in our models. The reason for choosing 10-fold method is that it could alleviate the overfitting and information leak. In all the optimizations, the number of trials was 100, and lightGBM hyperparameters were optimized by grid-search method.

Model Establishment

In order to get the best prediction model for CC, various combinations of clinicopathological characteristic and radiomic features were established using LightGBM. And then, 11 models were construct based on 1189 features, including a clinical model with only clinicopathological parameters (Model C); three radiomic model with radiomic features based on US alone, T2WI alone and DWI alone, respectively(Model R1, Model R2,Model R3,respectively ); two radiomic combined model (Model RR1 and Model RR2); and five combined model with the combination of clinicopathological characteristics and radiomic features(Model CR1, Model CR2, Model CR3, Model CR4 and Model CR5).

Model RR1: The Model RR1 consisted of radiomic features that were extracted from DWI and T2WI image.

Model RR2: The Model RR2 consisted of all radiomic features based on DWI, T2WI and US images.

Model CR1: The Model CR1 consisted of radiomic features based on US and clinicopathological parameters.

Model CR2: The Model CR2 consisted of radiomic features based on T2WI images and clinicopathological parameters.

Model CR3: The Model CR3 consisted of radiomic features based on DWI images and clinicopathological parameters.

Model CR4: The Model CR4 consisted of radiomic features based on DWI and T2WI images and clinicopathological parameters.

Model CR5: All of the clinicopathological and radiomic features formed the stable feature set.

Unsupervised clustering model

Features extracted from MRI and US images are of high dimensionality, and dimensionality reduction (DR) is an effective way to deal with these high dimensional features (Mwangi et al., 2014). Principal component analysis (PCA)(Jolliffe and Cadima, 2016) is a commonly used linear DR technique. PCA involves transforming the original data into a new coordinate system, where each axis represents a principal component, which is a linear combination of the original features. The motivation behind PCA is that it transforms high-dimensional data into a lower-dimensional space while preserving most of the important information. This makes it easier to visualize and analyze the data, and can improve the performance of machine learning algorithms. In addition, PCA can identify and remove correlated features, which can lead to more accurate and robust models.

We subsequently applied K-mean(MacQueen, 1967) clustering technique on the two scaling coordinates of PCA. K-mean clustering is a frequently-used unsupervised machine learning algorithm which can group similar data points into clusters. The algorithm works by iteratively partitioning a given dataset into K clusters, where K is a user-defined parameter representing the number of clusters to create.

Statistical analysis

SPSS (version. 22.0) were used to carry out the statistical analysis, and P < 0.05 was considered statistically significant. Machine learning algorithms were implemented by Python. Differences in the continuous parameters were evaluated using the Student’s T-test or Mann–Whitney U test, such as age and BMI. While chi-squared (χ2) test or Fisher’s exact test was used to compare the differences for the categorical variables, such as FIGO stage, tumor size and so on. Then, the maxstat package and survival package were applied for the optimal cut-point determination and Kaplan-Meier survival analysis.

Characteristic of the patients and Feature Selection

Over the study period, a total of 144 early-stage cervical cancer patients (average age, 58.76 ± 5.282 years) with all required data available that also met the inclusion criteria was finally included. The participants were then split into the training and testing cohort with the ratio of 6:4 randomly. Table 1 presented the subject baseline features of all patients involved in the training and testing cohort, and no statistically significant differences (P༞0.05)were found between cohorts in term of all characteristics.

The MRI and US images of cervical cancer for each patient were collected. Next 1180 imaging features with nonzero coefficients were included, among which 388 features were extracted from ultrasound, 396 from T2WI and 396 from DWI. The work flow of the model establishment was shown in Fig. 1.

Table 1

Patient characteristics in the training and testing cohort
Feature	Training(87)	Testing(57)	P value
Age	55.6897 ± 10.58225	56.7368 ± 10.17793	0.202
BMI^a	23.576459 ± 3.2704522	22.863823 ± 3.0842832	0.261
Tumor size	3.067 ± 1.3720	3.082 ± 1.0121	0.154
FIGO^b
Stage IB1	37	23	0.080
Stage IB2	5	4
Stage IIA1	27	24
Stage IIA2	18	6
LVSI^c			0.945
YES	27	18
NO	60	39
LNM^d
YES	20	6	0.075
NO	67	51
Differentiation
Low	2	1	0.063
Middle	36	21
High	49	35
DMI^e
༜1/3	16	9	0.093
༞1/3 ༜2/3	15	12
༞2/3	56	36
Pathology
Squamous cell carcinoma	79	53	0.095
Adenocarcinoma	3	3
Non adenocarcinoma and non squamous cell carcinoma	5	1

a body mass index

b International Federation of Gynecology and Obstetrics

c lymph-vascular space invasion

d lymph node status

e depth of myometrial invasion

Classification Supervised Machine Learning models and Evaluation

The most commonly used classification model in previous clinical research is logistic regression model. For comparison, we first established classification models using different machine learning algorithms incorporated all clinicopathological characteristics and radiomic features. The result presented that the LightGBM model achieved the best DFS prediction performance in testing cohort, and the corresponding values were 0.92 in accuracy and 0.86 in AUC. The accuracy of prediction and AUC were 0.83 and 0.72 with LRM shown in Fig. 2A; 0.91 and 0.59 with RF; 0.83 and 0.53 with NB; 0.88 and 0.65 with SVM; 0.59 and 0.68 with KNN, respectively.

Comparison of the models

According to Fig. 2A, the LightGBM model performed the best in DFS prediction, and thus it was used in subsequent predictive analysis for consistency. For further improve the performance of our model, various combinations of clinicopathological and radiomic features had been next tried in the study. Figure 2B showed the accuracy and ROC curve derived from multiple models for DFS prediction in the test cohort. The values for the highest predictive ACC and AUC curve among 11 models were 0.92 and 0.86 with Model CR5, respectively (red line). The result from Model CR5 was superior to those of any single or combined regression.

Next, the relative importance of a variable for DFS prediction was calculated. The top eight most factors for the model in order of importance were identified using LightGBM (Fig. 3C), including four radiomic features from ultrasound image and four from DWI images, but without any clinicopathological variables and T2WI radiomic features included. ShortRunLowGreyLevelEmphasis_angle90_offset from US was demonstrated as the critical variable among all features. Thereafter, standard box plots which presented the distribution of these top eight variables between relapse and non-recurrence was shown in Fig. 3B, showing the satisfactory results. It is well-known to render the model unstable due to multicollinearity once we use highly correlated features in regression analysis. Actually, there were moderate or weak correlations between theses top eight predictors according to Fig. 3A.

The detail top eight features were as follows:

ShortRunLowGreyLevelEmphasis_angle90_offset1.1 from US

LongRunEmphasis_angle45_offset7.2 from DWI

LowGreyLevelRunEmphasis_angle90_offset1.1 from US

MinIntensity.2 from DWI

ShortRunLowGreyLevelEmphasis_angle135_offset4.1 from US

ZonePercentage.2 from DWI

LowGreyLevelRunEmphasis_angle45_offset1.1 from US

Maximum3DDiameter.1 from DWI

Unsupervised clustering analysis using machine learning approach associated with prognosis

Next, unsupervised clustering which can visualize the similarity of the variables between any pair of samples, was used to identify CC patient subgroups related to DFS prognosis based on the all variables. Figure 4A illustrated a strong tendency toward the formation of two distinct groups in DFS rate among CC patients, and most of the patients with relapse were included in cluster 2 (blue cycle). Moreover, the Kaplan–Meier survival in Fig. 4B proved that DFS probability are statistically significantly differed across the two clusters (P = 0.025). In addition, clear differences in DFS rate between clusters were demonstrated in multiple features, and the top eight features were presented in Fig. 4C, which included four radiomic features from ultrasound image and four from DWI images.

In this hypothesis generating study, we investigated the performance of 11 models based on clinicopathological variables and radiomic features from MRI and ultrasound for the capability towards DFS prediction of early-stage cervical cancer. LightGBM outperforms the other machine learning algorithms, with radiomic feature identified as the most effective predictor. ROC analysis proved that model CR5 with all variables included showed the best performance, yielding with AUC of 0.86 and ACC of 0.92. In general, this study provides new evidence for the combination of clinicopathological and radiomic features based on LightGBM, which may help promote the individualized post-treatment strategy plan that clinicians can follow, from active monitoring to chemotherapy and radiotherapy.

MRI imaging is commonly used to access tumor size, extent of invasion into adjacent structures and lymph nodes metastatic in cervical cancer, and surgery is then considered when the tumor is confined to the cervix/upper vagina. The role of radiomic from MRI, and its possible applications in CC, has been paid more attention recently. Cai et al (Cai et al., 2021) established a radiomics model based on T2WI and DWI ,and demonstrated that the model had a good predictive ability in progression-free survival, which yielded a C-index of 0.803 in the training cohort and an AUC of 0.795 in the validation cohort. Fang et al (Fang et al., 2020) similarly confirmed the value of radiomics features from MRI in DFS prediction by analyzing the values of selected features involving 248 patients with early-stage cervical cancer.

Technical development of US has currently led to its application in CC. Belitsos et al(Belitsos et al., 2012) proved the accuracy of ultrasound in differentiating precancerous lesion from cervical cancer, which improved the diagnostic performance. Jin et al (Jin et al., 2020) constructed a radiomic prediction model based on US images to assess predictive power of the LMN in early-stage CC, which showed satisfactory ability of US to discriminate LNM status. Furthermore, Fischerova et al (Fischerova et al., 2008) compared the application of US with MRI in the identification of tumor size. The accuracy of US was 90.5% whereas of MRI was 81.1% (P < or = 0.049), confirming the utility of US. It is worth noting that the combination of technique might play a significant role in tumor prediction (Millischer et al., 2015; Theodore et al., 2017). The investigators (Millischer et al., 2015; Moro et al., 2020; Theodore et al., 2017) showed that MRI and US could improve CC characterization and parametrial infiltration detection when compared with MRI alone. In addition, further evidence regarding the capability of technique combination for cervical cancer diagnosis has been suggested by Moro et al (Moro et al., 2020).

However, with the development of tumor research, there are more and more variables from different aspects, such as radiation, laboratory, pathology and genetics. Thus, Cai et al (Cai et al., 2021) found that the superiority of the combined model over radiomic feature or clinical features alone to CC survival prediction. This finding was then reproduced by Zheng et al (Zheng et al., 2022)and Zhang et al(Zhang et al., 2022). In these works, they also demonstrated that the model would perform better when these radiomic and clinical variables were incorporated. Our study is concordant with this in showing the predictive model with all variables incorporated has the best performance. In our study, we calculated the relative importance of a variable for DFS prediction, and found that the top eight factors of importance all related to radiomic features, including 4 features from US images and 4 from DWI images. The same result was found after unsupervised clustering analysis applied. Interesting, none of clinicopathological variables were found to participate in these selected subsets. We consider this an important discovery, consistent with previous studies(Cai et al., 2021; Zheng et al., 2022), where radiomic feature were proved to have more impact on performance than clinical features. The outstanding performance of radiomic features indicates that the information hidden in the images might reveal disease features and reflect underlying pathophysiology. Importantly, radiomics would yield promising results in previous studies, such as survival, tumor progression or genetic mutations (Hou et al., 2020; Wang et al., 2020a; Yu et al., 2020; Zhang et al., 2022), when together with ML models that are data-driven analysis methods for mining implicit clinical values from image features.

To date, the most commonly used classification model in clinical research is logistic regression model. Although good results have been proved, the application were not so extensive. As a generalized linear model, logistic model can better deal with the linear relationship between variables. Our study is line with this in showing the good performance of logistic model in DFS prediction. However, the fact is that there is not only a linear correlation but also a non-linear relationship among the characteristics. Thus, LightGBM shows a better performance in this study when compared with the logistic model, reflecting the reliability of our study to some extent. Recently, LightGBM is a newly developed ML model, which could tune a larger number of variables to ensure the accuracy of the model, achieving the optimal performance. While this algorithm has been successfully used in other cancer researches recently (Dong et al., 2020; Vamvakas et al., 2022), but not yet in cervical cancer. Our study demonstrated that the application of LightGBM model achieved an average AUC of 0.86, overcoming the limited efficacy of traditional ML models in improving the CC diagnostic accuracy.

Since the supervised machine learning approach with all variables is proved as an accurate prognostic prediction tool for the management of CC, we additionally implemented steps of unsupervised clustering analysis for prognosis analysis. Interestingly, unsupervised clustering analysis based on all variables was also found to effectively segregate CC subgroups which were clearly associated with prognosis. And a recent analysis of study (Kawakami et al., 2019) included 334 patients with epithelial ovarian cancer showed unsupervised clustering analysis could identify subgroups with significantly worse survival.

While the study highlights the potential of the integration of radiomic and LightGBM, some limitations of this study still deserved to be reported. Specifically, the selected sample size was small due to patients without MRI or US images in this hospital were excluded, which might inevitably limit the performance of the model in our research. Going forward, our study was performed without internal and external validation, which have a certain impact on the reduction of model accuracy confirmation intensity. All in all, a more comprehensive model with larger sample and multicenter external validation was looking forward to establishing in future.

In summary, our study demonstrated that the integration of MRI and US based radiomics features with clinicopathological variables has the potential to predict DFS in early-stage CC patients with LightGBM classifier. Nevertheless, further studies are need to assess the performance of radiomic and LightGBM classifier in cervical cancer.

Statements and Declarations

Funding

This work was supported by Wenzhou Association for Science and Technology(grant No:kjfw103).

Competing Interests

The authors have no relevant financial or non-financial interests to disclose.

Author Contributions

All authors contributed to the study conception and design. XWH conceived and designed the analysis, collected the data, wrote the paper. ZLR and WTX collected the data and performed the analysis. XQF identified eligible patients through the Pathology database. JYM , FL aided in designing the analysis, and Martin Powell aided in editing the paper. CJ conceived and designed the analysis, contributed data and editing the paper.

Data Availability Statement

The data underlying this article will be shared on reasonable request to the corresponding author.

Ethics approval

This retrospective study was approved by the Ethics Committee of the First Affiliated Hospital of Wenzhou Medical University(number: 2021-zz-048). We confirmed that all methods were performed in accordance with the relevant guidelines and regulations (e.g., Declaration of Helsinki).

Consent to participate

The requirement for informed consent was waived by the Ethics Committee of the First Affiliated Hospital of Wenzhou Medical University because of the retrospective nature of the study.

Ayhan A., Al R.A., Baykal C., Demirtas E., Ayhan A., Yüce K. (2004) Prognostic factors in FIGO stage IB cervical cancer without lymph node metastasis and the role of adjuvant radiotherapy after radical hysterectomy. Int J Gynecol Cancer 14:286–92. DOI: 10.1111/j.1048-891X.2004.014212.x.
Balcacer P., Shergill A., Litkouhi B. (2019) MRI of cervical cancer with a surgical perspective: staging, prognostic implications and pitfalls. Abdom Radiol (NY) 44:2557–2571. DOI: 10.1007/s00261-019-01984-7.
Barter J.F., Soong S.J., Shingleton H.M., Hatch K.D., Orr J.W., Jr. (1989) Complications of combined radical hysterectomy-postoperative radiation therapy in women with early stage cervical cancer. Gynecol Oncol 32:292–6. DOI: 10.1016/0090-8258(89)90627-6.
Belitsos P., Papoutsis D., Rodolakis A., Mesogitis S., Antsaklis A. (2012) Three-dimensional power Doppler ultrasound for the study of cervical cancer and precancerous lesions. Ultrasound Obstet Gynecol 40:576–81. DOI: 10.1002/uog.11134.
Berger J., Henneman O., Rhemrev J., Smeets M., Jansen F.W. (2018) MRI-Ultrasound Fusion Imaging for Diagnosis of Deep Infiltrating Endometriosis - A Critical Appraisal. Ultrasound Int Open 4:E85-e90. DOI: 10.1055/a-0647-1575.
Bhatla N., Berek J.S., Cuello Fredes M., Denny L.A., Grenman S., Karunaratne K., Kehoe S.T., Konishi I., Olawaiye A.B., Prat J., Sankaranarayanan R., Brierley J., Mutch D., Querleu D., Cibula D., Quinn M., Botha H., Sigurd L., Rice L., Ryu H.S., Ngan H., Mäenpää J., Andrijono A., Purwoto G., Maheshwari A., Bafna U.D., Plante M., Natarajan J. (2019) Revised FIGO staging for carcinoma of the cervix uteri. Int J Gynaecol Obstet 145:129–135. DOI: 10.1002/ijgo.12749.
Cai M., Yao F., Ding J., Zheng R., Huang X., Yang Y., Lin F., Hu Z. (2021) MRI Radiomic Features: A Potential Biomarker for Progression-Free Survival Prediction of Patients With Locally Advanced Cervical Cancer Undergoing Surgery. Frontiers in Oncology 11. DOI: 10.3389/fonc.2021.749114.
Castiglioni I., Gallivanone F., Soda P., Avanzo M., Stancanello J., Aiello M., Interlenghi M., Salvatore M. (2019) AI-based applications in hybrid imaging: how to build smart and truly multi-parametric decision models for radiomics. Eur J Nucl Med Mol Imaging 46:2673–2699. DOI: 10.1007/s00259-019-04414-4.
Dias R., Torkamani A. (2019) Artificial intelligence in clinical and genomic diagnostics. Genome Med 11:70. DOI: 10.1186/s13073-019-0689-8.
Dong X., Dan X., Yawen A., Haibo X., Huan L., Mengqi T., Linglong C., Zhao R. (2020) Identifying sarcopenia in advanced non-small cell lung cancer patients using skeletal muscle CT radiomics and machine learning. Thorac Cancer 11:2650–2659. DOI: 10.1111/1759-7714.13598.
Fang J., Zhang B., Wang S., Jin Y., Wang F., Ding Y., Chen Q., Chen L., Li Y., Li M., Chen Z., Liu L., Liu Z., Tian J., Zhang S. (2020) Association of MRI-derived radiomic biomarker with disease-free survival in patients with early-stage cervical cancer. Theranostics 10:2284–2292. DOI: 10.7150/thno.37429.
Fischerova D., Cibula D., Stenhova H., Vondrichova H., Calda P., Zikan M., Freitag P., Slama J., Dundr P., Belacek J. (2008) Transrectal ultrasound and magnetic resonance imaging in staging of early cervical cancer. Int J Gynecol Cancer 18:766–72. DOI: 10.1111/j.1525-1438.2007.01072.x.
Forghani R., Savadjiev P., Chatterjee A., Muthukrishnan N., Reinhold C., Forghani B. (2019) Radiomics and Artificial Intelligence for Biomarker and Prediction Model Development in Oncology. Comput Struct Biotechnol J 17:995–1008. DOI: 10.1016/j.csbj.2019.07.001.
Haldorsen I.S., Lura N., Blaakær J., Fischerova D., Werner H.M.J. (2019) What Is the Role of Imaging at Primary Diagnostic Work-Up in Uterine Cervical Cancer? Curr Oncol Rep 21:77. DOI: 10.1007/s11912-019-0824-0.
Hou L., Zhou W., Ren J., Du X., Xin L., Zhao X., Cui Y., Zhang R. (2020) Radiomics Analysis of Multiparametric MRI for the Preoperative Prediction of Lymph Node Metastasis in Cervical Cancer. Front Oncol 10:1393. DOI: 10.3389/fonc.2020.01393.
Huang H., Liu Q., Zhu L., Zhang Y., Lu X., Wu Y., Liu L. (2019) Prognostic Value of Preoperative Systemic Immune-Inflammation Index in Patients with Cervical Cancer. Scientific reports 9:3284. DOI: 10.1038/s41598-019-39150-0.
Jin X., Ai Y., Zhang J., Zhu H., Jin J., Teng Y., Chen B., Xie C. (2020) Noninvasive prediction of lymph node status for patients with early-stage cervical cancer based on radiomics features from ultrasound images. Eur Radiol 30:4117–4124. DOI: 10.1007/s00330-020-06692-1.
Jolliffe I.T., Cadima J. (2016) Principal component analysis: a review and recent developments. Philos Trans A Math Phys Eng Sci 374:20150202. DOI: 10.1098/rsta.2015.0202.
Kawakami E., Tabata J., Yanaihara N., Ishikawa T., Koseki K., Iida Y., Saito M., Komazaki H., Shapiro J.S., Goto C., Akiyama Y., Saito R., Saito M., Takano H., Yamada K., Okamoto A. (2019) Application of Artificial Intelligence for Preoperative Diagnostic and Prognostic Prediction in Epithelial Ovarian Cancer Based on Blood Biomarkers. Clin Cancer Res 25:3006–3015. DOI: 10.1158/1078-0432.Ccr-18-3378.
Kobayashi Y., Yoshida K. (2021) Quantitative structure-property relationships for the calculation of the soil adsorption coefficient using machine learning algorithms with calculated chemical properties from open-source software. Environ Res 196:110363. DOI: 10.1016/j.envres.2020.110363.
Koh W., Abu-Rustum N., Bean S., Bradley K., Campos S., Cho K., Chon H., Chu C., Clark R., Cohn D., Crispens M., Damast S., Dorigo O., Eifel P., Fisher C., Frederick P., Gaffney D., Han E., Huh W., Lurain J., Mariani A., Mutch D., Nagel C., Nekhlyudov L., Fader A., Remmenga S., Reynolds R., Tillmanns T., Ueda S., Wyse E., Yashar C., McMillian N., Scavone J. (2019) Cervical Cancer, Version 3.2019, NCCN Clinical Practice Guidelines in Oncology. Journal of the National Comprehensive Cancer Network: JNCCN 17:64–84. DOI: 10.6004/jnccn.2019.0001.
Lu A., Lu G. (2022) Application of MRI and CT Images in Surgical Treatment of Early Cervical Cancer. Scanning 2022:1592449. DOI: 10.1155/2022/1592449.
MacQueen J. (1967) Some methods for classification and analysis of multivariate observations. Proceedings of the Fifth Berkeley Symposium on Mathematical Statistics and Probability:281–297.
Maguire R., Kotronoulas G., Simpson M., Paterson C. (2015) A systematic review of the supportive care needs of women living with and beyond cervical cancer. Gynecol Oncol 136:478–90. DOI: 10.1016/j.ygyno.2014.10.030.
Manganaro L., Lakhman Y., Bharwani N., Gui B., Gigli S., Vinci V., Rizzo S., Kido A., Cunha T.M., Sala E., Rockall A., Forstner R., Nougaret S. (2021) Staging, recurrence and follow-up of uterine cervical cancer using MRI: Updated Guidelines of the European Society of Urogenital Radiology after revised FIGO staging 2018. Eur Radiol 31:7802–7816. DOI: 10.1007/s00330-020-07632-9.
Merz J., Bossart M., Bamberg F., Eisenblaetter M. (2020) Revised FIGO Staging for Cervical Cancer - A New Role for MRI. Rofo 192:937–944. DOI: 10.1055/a-1198-5729.
Millischer A.E., Salomon L.J., Santulli P., Borghese B., Dousset B., Chapron C. (2015) Fusion imaging for evaluation of deep infiltrating endometriosis: feasibility and preliminary results. Ultrasound Obstet Gynecol 46:109–17. DOI: 10.1002/uog.14712.
Moro F., Gui B., Arciuolo D., Bertoldo V., Borzi R., Romeo P., Petta F., Cambi F., Pasciuto T., Zannoni G.F., Valentini V., Manfredi R., Scambia G., Testa A.C. (2020) Fusion imaging of ultrasound and MRI in the assessment of locally advanced cervical cancer: a prospective study. Int J Gynecol Cancer 30:456–465. DOI: 10.1136/ijgc-2019-000902.
Mwangi B., Tian T.S., Soares J.C. (2014) A review of feature reduction techniques in neuroimaging. Neuroinformatics 12:229–44. DOI: 10.1007/s12021-013-9204-3.
Papathemelis T., Knobloch S., Gerken M., Scharl A., Anapolski M., Ignatov A., Inwald E., Ortmann O., Scharl S., Klinkhammer-Schalke M. (2019) Impact of nodal status and treatment strategy on overall survival in advanced stage cervical cancer. Journal of cancer research and clinical oncology 145:1369–1376. DOI: 10.1007/s00432-019-02890-7.
Siegel R.L., Miller K.D., Fuchs H.E., Jemal A. (2022) Cancer statistics, 2022. CA Cancer J Clin 72:7–33. DOI: 10.3322/caac.21708.
Theodore C., Levaillant J.M., Capmas P., Chabi N., Skalli D., Vienet-Legue L., Haddad B., Fernandez H., Touboul C. (2017) MRI and Ultrasound Fusion Imaging for Cervical Cancer. Anticancer Res 37:5079–5085. DOI: 10.21873/anticanres.11925.
Tian Y., Luo H. (2022) Diagnostic accuracy of transvaginal ultrasound examination for local staging of cervical cancer: a systematic review and meta-analysis. Med Ultrason 24:348–355. DOI: 10.11152/mu-3246.
Vamvakas A., Tsivaka D., Logothetis A., Vassiou K., Tsougos I. (2022) Breast Cancer Classification on Multiparametric MRI - Increased Performance of Boosting Ensemble Methods. Technol Cancer Res Treat 21:15330338221087828. DOI: 10.1177/15330338221087828.
Wang T., Gao T., Guo H., Wang Y., Zhou X., Tian J., Huang L., Zhang M. (2020a) Preoperative prediction of parametrial invasion in early-stage cervical cancer with MRI-based radiomics nomogram. Eur Radiol 30:3585–3593. DOI: 10.1007/s00330-019-06655-1.
Wang Y., Zhang Z., Tao P., Reyila M., Qi X., Yang J. (2020b) The Abnormal Expression of miR-205-5p, miR-195-5p, and VEGF-A in Human Cervical Cancer Is Related to the Treatment of Venous Thromboembolism. Biomed Res Int 2020:3929435. DOI: 10.1155/2020/3929435.
Yu S., Li X., Zhang J., Wu S. (2021) Development of a Novel Immune Infiltration-Based Gene Signature to Predict Prognosis and Immunotherapy Response of Patients With Cervical Cancer. Front Immunol 12:709493. DOI: 10.3389/fimmu.2021.709493.
Yu Y., Tan Y., Xie C., Hu Q., Ouyang J., Chen Y., Gu Y., Li A., Lu N., He Z., Yang Y., Chen K., Ma J., Li C., Ma M., Li X., Zhang R., Zhong H., Ou Q., Zhang Y., He Y., Li G., Wu Z., Su F., Song E., Yao H. (2020) Development and Validation of a Preoperative Magnetic Resonance Imaging Radiomics-Based Signature to Predict Axillary Lymph Node Metastasis and Disease-Free Survival in Patients With Early-Stage Breast Cancer. JAMA Netw Open 3:e2028086. DOI: 10.1001/jamanetworkopen.2020.28086.
Zhang X., Zhao J., Zhang Q., Wang S., Zhang J., An J., Xie L., Yu X., Zhao X. (2022) MRI-based radiomics value for predicting the survival of patients with locally advanced cervical squamous cell cancer treated with concurrent chemoradiotherapy. Cancer Imaging 22:35. DOI: 10.1186/s40644-022-00474-2.
Zheng R., Cai M., Lan L., Huang X., Yang Y., Powell M., Lin F. (2022) An MRI-based radiomics signature and clinical characteristics for survival prediction in early-stage cervical cancer. The British journal of radiology 95:20210838. DOI: 10.1259/bjr.20210838.

No competing interests reported.

SupplementaryMethods1.doc

Download PDF

Version 1

posted

You are reading this latest preprint version

Application of Machine Learning for Prognostic Prediction in early-stage Cervical Cancer, Based on radiomics from ultrasound and MRI

Status:

Version 1

Abstract

Figures

Introduction

Materials and methods

Patient data

Retrieval of Clinicopathological information

Image acquisition, tumor segmentation and radiomics

Image acquisition and tumor segmentation

Radiomics feature extraction

Supervised Machine Learning Classifiers

LightGBM

Model Establishment

Unsupervised clustering model

Statistical analysis

Results

Characteristic of the patients and Feature Selection

Classification Supervised Machine Learning models and Evaluation

Comparison of the models

Unsupervised clustering analysis using machine learning approach associated with prognosis

Discussion

Conclusion

Declarations

References

Additional Declarations

Supplementary Files

Status:

Version 1