Radiomics based predictive modeling of rectal toxicity in prostate cancer patients undergoing radiotherapy: CT and MRI comparison

Rectal toxicity is one of the common side effects after radiotherapy in prostate cancer patients. Radiomics is a non-invasive and low-cost method for developing models of predicting radiation toxicity that does not have the limitations of previous methods. These models have been developed using individual patients’ information and have reliable and acceptable performance. This study was conducted by evaluating the radiomic features of computed tomography (CT) and magnetic resonance (MR) images and using machine learning (ML) methods to predict radiation-induced rectal toxicity. Seventy men with pathologically confirmed prostate cancer, eligible for three-dimensional radiation therapy (3DCRT) participated in this prospective trial. Rectal wall CT and MR images were used to extract first-order, shape-based, and textural features. The least absolute shrinkage and selection operator (LASSO) was used for feature selection. Classifiers such as Random Forest (RF), Decision Tree (DT), Logistic Regression (LR), and K-Nearest Neighbors (KNN) were used to create models based on radiomic, dosimetric, and clinical data alone or in combination. The area under the curve (AUC) of the receiver operating characteristic curve (ROC), accuracy, sensitivity, and specificity were used to assess each model’s performance. The best outcomes were achieved by the radiomic features of MR images in conjunction with clinical and dosimetric data, with a mean of AUC: 0.79, accuracy: 77.75%, specificity: 82.15%, and sensitivity: 67%. This research showed that as radiomic signatures for predicting radiation-induced rectal toxicity, MR images outperform CT images.


Introduction
One of the most prevalent cancers in males is prostate cancer.Radiotherapy was introduced as an accepted curative treatment method for prostate cancer.One of the common side effects of this treatment is exposure to radiation in healthy organs close to the prostate like the rectum [1,2].This unavoidable radiation has several immediate and delayed impacts on the rectum, and the probability and intensity of this toxicity depend on several clinical factors [3,4].Radiation toxicity is an effective factor to limit the therapeutic dose.A wide range of rectal toxicity was observed in prostate cancer patients undergoing radiotherapy in prior clinical experiences [5,6].
To enhance the effectiveness of therapy and the quality of life of patients, predictive modeling of treatmentinduced toxicity rates might be helpful [7][8][9].In addition to informing the patient about the possible side effects of treatment, these modelings can obtain the possibility for clinicians to adjust the treatment based on the patient's characteristics to decrease the risk of side effects.There are several studies to model these complications which have limitations, such as differences in the radiation sensitivity of patients, differences in the radiation sensitivity of different organs, uncertainty in dosimetric parameters, and programming of models [10,11].One of the unmet needs in modern radiation therapy is the lack of a prognostic biomarker for radiation-related morbidity.Therefore, it is necessary to use additional patient-specific and repeatable factors to improve the prediction models.
Integrating imaging biomarkers into clinical and dosimetric data has resulted in predictive models with improved performance because of the developments in imaging biomarker research, including radiomic studies.Modeling based on radiomic features, in addition to being a simple, low-cost, and non-invasive method, is based on individual patient information, it can provide reliable prediction models with reproducibility and acceptable performance [12,13].With the advent of radiomics along with machine learning and deep learning methods, quantitative image parameters can be extracted and used for radiotherapy applications.Radiomics has been used as an advanced image processing model to extract quantitative features from medical images that can identify important phenotypic variations in malignancies and be used in the fields of diagnosis, prognosis, and treatment with the help of machine learning and deep learning methods [14][15][16].
Using quantitative imaging, the limitations associated with molecular profiling can be overcome in personalized oncology [17,18].Hence, several applications of radiomics in the field of personalized oncology include stage distinction [19], molecular grade [20][21][22], the prognostic effect [23], radiation toxicity assessment [24], and treatment response prediction [25,26].According to research, it is feasible to improve the prediction of radiation proctitis in organs such as the rectum [27,28], bladder [28,29], lung [30], and ear [31] by using features derived from medical pictures coupled with clinical and dosimetric information.
In the current work, we developed prediction models for radiotherapy-induced rectal toxicity using radiomic analysis on magnetic resonance imaging (MRI) and computed tomography (CT) images of the rectal wall before treatment.These models are a non-invasive and low-cost method and do not have the limitations of the past methods, such as the difference in the radiation sensitivity of different organs and the difference in the radiation sensitivity of the patient.Also, because these models are developed using individual information of patients, including quantitative features of medical images and clinical and dosimetric parameters of patients, they are reliable, repeatable, and with acceptable performance.This study's primary objective was to assess the effect of radiomics features of pretreatment CT and MR images to identify radiation-induced rectal toxicity in prostate radiation therapy.as well as to test the idea that MR images' radiomic properties are more effective than CT images as radiomic signatures for predicting radiationinduced rectal toxicity.To the best of our knowledge, this is the first study that compares prediction models of rectal radiation toxicity developed based on radiomic features of CT and MR images.

Patients
Between January 2021 and April 2022, this prospective research was carried out with the agreement of the local ethics committee (Isfahan University of Medical Sciences, Isfahan, Iran) with the ethics approval number (IR.MUI.MED.REC.1399.731).In this research, 70 prostate cancer patients with a pathology-confirmed diagnosis had examinations after providing written informed consent.This investigation was following the Helsinki Declaration.Our inclusion criteria were; (a) The adenocarcinoma of the prostate was histologically confirmed, (b) The external radiation therapy candidate, (c) Eastern Cooperative Oncology Group (ECOG) Performance Status 0-2.Also, our exclusion criteria were: (a) Patients with metastatic prostate cancer, (b) Patients who have had any prostate surgery, (c) Patients with kidney and liver disorders, and (d) gastrointestinal dysfunction such as reflux.

Image dataset
Before beginning radiotherapy, CT and MRI scans were done on each patient.The same CT scanner was used for all CT scans (Siemens, 16 slices).KV = 110, mAs = 225, field of view (FOV) = 380 × 380 mm, matrix = 512 × 512, thickness = 3.0 mm, and gap = 0.8 mm were used to create the CT images.Additionally, MRI scans were carried out at a medical imaging facility using 1.5 tesla Siemens MRI equipment.All patients had the same imaging procedures, including diffusion-weighted imaging (DWI) and T2-weighted imaging (T2W).Using the b value shown in Table 1, the imager software automatically determined the apparent diffusion coefficient (ADC).

Radiotherapy and toxicity assessment
Three-Dimensional Radiation Therapy (3DCRT) was administered to all patients in four fields at a dosage of 45 Gy over 25 fractions.Using the Prowess Panther (v.5.5)Treatment Planning System (TPS), 3D-CRT plans were created.During therapy, radiation-induced toxicity was evaluated once a week in accordance with the Common Terminology Criteria for Adverse Events (CTCAE v5.0).Radiation-induced ≥ grade 1 proctitis was considered as the main rectal toxicity.

Feature extraction
Pre-processed images were used for feature extraction.To process, images were discretized and sampled into 10 Gy bins, and a Laplacian Gaussian (LOG) filter with a sigma value of 0.5 was employed for pre-processing.After using pre-processing filters, one hundred features were extracted from every patient's CT and MR images utilizing the opensource software package 3D slicer v.4.11.The features included first-order features, shape-based and textural features.The texture sets contained the neighbor gray-tone difference matrix (NGTDM), the gray-level run length matrix (GLRLM), the gray-level co-occurrence matrix (GLCM), the gray-level size zone matrix (GLSZM), and the graylevel dependency matrix (GLDM).The key feature value was finally determined to be the average feature value.Radiomic features extracted from the different images were detailed in Supplementary Table 1.

Feature selection
It is important to select the features extracted from images for radiomic analysis.Many of them are just noise or are highly correlated with each other.Improper selection of features can increase the probability of overfitting, increase the computational cost and decrease the prediction accuracy.For assessing the predictive capacity of the radiomic features of CT and MR images and clinical and dosimetric parameters, the receiver operating characteristic (ROC) curve was computed.Then, the least absolute shrinkage and selection operator (LASSO) method was used to select features with non-zero coefficients that are useful for prediction.5 subgroups.Of these 5 subsets, each time one was used for validation (test data) and the other 4 were used for training (training data).This process was repeated 5 times and all data were used exactly once for training and once for validation.Finally, the average result of these 5 times validation was chosen as the final estimate.The area under the curve (AUC) of the receiver operating characteristic curve (ROC), accuracy, sensitivity, and specificity parameters was then used to determine the capacity of the aforementioned models to predict radiation toxicity.Usually, in ML modeling, the standard cut-off value is considered 50%.The accuracy, sensitivity, and specificity of models are calculated using the given formulas below [32]: Where ''TN'' stands for True Negative, which indicates the number of negative samples that are accurately classified in the confusion matrix.Similarly, ''TP'' stands for True Positive, which indicates the number of positive examples that are accurately classified in the confusion matrix.The term "FP" indicates the False Positive value in the confusion matrix, i.e., the number of true negative samples classified as positive.And "FN" stands for False Negative value in the confusion matrix, which is the number of true positive samples classified as negative.

Patient
The study included 70 men with prostate cancer, ranging in age from 51 to 93, with a mean age of 71.5 years.39 (56%) of the 70 individuals experienced radiation proctitis.Table 2 has further information about the patients.

Dosimetric and clinical parameter
For every patient, the dosimetric parameters, including Equivalent uniform dose (EUD), mean/min/max dose, D5-D100 and V5-V100, and total volume for the rectal wall were extracted from the TPS.The clinical parameters, including age, PSA level (PSA level before and after radiotherapy), gleason score, stage, underlying disease, and addiction were provided for all patients.

Classifier model
In this work, two patient groups with and without proctitis were compared using binary classification and several machine learning (ML) techniques.As a result, individuals with proctitis were classified as class 0 and those without proctitis as class 1. Predictive models were created using radiomic, dosimetric, and clinical variables alone or in combination.To construct a prediction model, selected characteristics with the LASSO method were employed as covariates (X), and the findings of rectal toxicity as the dependent variable (Y) in the Random Forest (RF), Decision Tree (DT), Logistic Regression (LR), and K-nearest neighbors (KNN) algorithms.Using the grid-search cross validation method, the optimal parameters of these algorithms were determined.ML algorithms were applied on the Python package "scikit-learn" (version 0.24.2).To determine the validity of prediction models and to achieve the best results, and also because the number of patient samples was small, the 5-fold cross-validation method was used.In this type of validation, the data were randomly divided into with clinical and dosimetric features with an AUC of 0.84 (accuracy: 81%; sensitivity: 74%; and specificity: 87%).
The receiver operating characteristic (ROC) curve of KNN models based on radiomic aspects of CT and MR images alone or in conjunction with dosimetric and clinical data is shown in Fig. 4.This image shows that the performance of the models is enhanced by integrating the radiomic features of images with clinical and dosimetric data.
The results of all models for predicting radiation toxicity of the rectum based on AUC, Accuracy (Ac), Sensitivity (Se), and Specificity (Sp) parameters are shown in Table 3. characteristics chosen (using the LASSO approach) as the primary features in various feature sets.

Prediction results
Clinical and dosimetric predictors were added to chosen distinct radiomic features, and then ML models were used to create prediction models.Figure 3 displays a heat map of the AUC, accuracy, sensitivity, and specificity for various model and feature set combinations.According to this Figure, the KNN classifier provided the best capability to predict radiation-induced rectal toxicity based on radiomic features of MR images in combination with dosimetric features with an AUC of 0.86 (accuracy: 79%; sensitivity: 63%; and specificity: 91%).Moreover, the RF classifier showed an excellent performance in predicting radiation proctitis based on radiomic features of MR images in combination The present study identified important features, including clinical, dosimetry, shape, and texture to predict radiationinduced rectal toxicity.Features from shape and texture families include Coarseness, GrayLevelNonUniformity, Maximum2DdiameterColumn, SmallDependenceHigh-GrayLevelEmphasis, MCC, Contrast, and Flatness.These features are the measures of tissue heterogeneity or homogeneity.Moreover, because of the higher soft tissue contrast of MR images compared to CT, the importance of extracting features from MR images to predict rectal toxicity is higher.
In line with the present study, Abdullahi et al. [27] using radiomic features extracted from pre and post-treatment MRI of the rectal wall, developed predictive radiomic models and evaluated their performance.In their study, some features, such as Non-Uniformity and Short Run Emphasis which belong to the GLRLM family, were found as significant features associated with radiation toxicity.These features are indicators of tissue non-uniformity or heterogeneity.In another study, Shayan Mostafaei et al. [28] created prediction models for radiation cystitis and proctitis in prostate cancer patients based on radiomic features of CT images and clinical and dosimetry parameters.In their research, radiomic features including GLRLM, GLSZM, and GLDM

Discussion
One of the side effects caused by radiation in prostate cancer patients treated with radiotherapy is rectal toxicity.To predict these side effects, various clinical and dosimetric parameters have limitations.Advances in imaging biomarker research, including radiomic studies and their combination with clinical and dosimetric parameters, have led to the development of predictive models with better reproducibility and performance.
This study compared the radiomic features of pre-treatment rectal wall CT and MR images to predict radiationinduced rectal toxicity.The results showed that the radiomic features of MR images have a better performance compared to CT images as radiomic signatures to predict radiationinduced rectal toxicity.Additionally, some radiomic models were developed in this research using clinical, dosimetric, and imaging data.By adding clinical and dosimetric factors to created models based on the radiomic features of CT and MR images, radiation toxicity prediction of rectal may be greatly improved.The models based on the radiomic features of MR images perform better than those based on the radiomic features of CT images in terms of AUC, Accuracy, Sensitivity, and Specificity parameters, as shown in Fig. 5.The greater contrast of soft tissue in MR images has improved the accuracy of the models' rectal toxicity prediction.On the other hand, combining patients' clinical and dosimetric information with radiomic features increases the performance of all models.The collected findings demonstrate that among the four alternative ML models examined, KNN and RF models performed the best, with AUCs of 0.86 and 0.84 and accuracy of 0.79 and 0.81, respectively.Therefore, these models can be applied to predict rectal toxicity.
Several researchers have shown the use of ML models to assess images from several modalities and find the optimal model.For instance, Felix Peisen et al. [33] presented different results from the results of the present study.They investigated models.In the test dataset, the AUCs for the clinical, LDA, MLP, and SGD models were 0.517, 0.719, 0.704, and 0.725, respectively.According to their findings, the prediction model that included clinical and radiomic features and had an AUC of 0.754 in the dataset of the test performed the best.Also, Qingying Yang et al. [35] constructed an ideal radiomic model utilizing non-contrast CT for PH prediction based on radiomic features and compared their performance with prediction models based on clinical and radiological factors using ten different ML models.Their results showed that the support vector machine (SVM) model had the highest prediction performance with an AUC of 0.879 and an accuracy of 0.834.Also, they revealed that the combined predictive model had the best performance based on radiomic features along with clinical and radiological parameters.Additionally, Sepideh Amiri et al. [36] developed and investigated prediction models based on the radiomic characteristics of computed tomography images to predict the risk of chronic kidney disease.In this study, they created predictive models using radiomic characteristics of CT images of 50 patients with abdominal cancers who underwent radiation therapy and naive bayesian bernoulli, decision tree, gradient boosting decision trees, K-nearest neighbor, random forest, and support vector machine algorithms.Their results showed that a combination of radiomic and clinical parameters could predict radiotherapy-induced chronic kidney toxicity.In another study, Feng Du et al. [37] developed and investigated a novel nomogram based on the radiomic features of cone beam computed tomography (CBCT) images in different periods during radiation therapy (RT) and combined them with clinical and dosimetric parameters to predict radiation pneumonitis (RP).In this retrospective study, they developed predictive models using radiomic features of CBCT images of 96 patients with esophageal squamous cell carcinoma (ESCC) and a logistic regression machine learning algorithm.According to their results, the nomogram model developed based on the radiomic features of CBCT images and their combination with clinical and dosimetric parameters has a better predictive ability than other prediction models and can be used as a predictive model for RP.They showed that the use of radiomic features and their combination with clinical parameters improved the performance of predictive models.Also, the table of recent papers was added in supplementary Table 2.These differences can have various reasons, including images with different modalities, devices, and parameters of image acquisition, image segmentation, zone of the tumor, and feature selection methods [38].Furthermore, ML methods are not superior to each other and there is no specific and stable ML algorithm for all problems.various methods.Also, to increase the reliability of the developed models, an external data set can be used to validate the models.Since the ML methods are not superior to the comparison and there is no specific and fixed ML for all problems, therefore a combination of selectors and classifiers will result in different performances and results.In this study, only one selector and four classifiers were investigated.Therefore, it is suggested to investigate a

Limitations
The primary constraint on this research was data size.Consequently, 5-fold cross-validation was used to verify the models.In addition to circumventing the size constraint, this decreased the findings' sensitivity to the input data and improved dependability.Therefore, it is recommended that larger data sets be employed in future studies to compare  combination of different selectors and classifiers and compare their performance in future studies.Also, in this study, machine learning methods were used to develop prediction models of rectal radiation toxicity and for future studies, deep learning methods can be used and their performance can be compared with machine learning methods.In this study, to compare the ability of radiomic features of MRI and CT images to develop prediction models of rectal radiation toxicity, fusing MRI and CT images was not done, which can be done in future studies.

Conclusions
One of the limitations of prostate cancer treated with radiation therapy is the side effects of this method, including rectal radiation toxicity.To reduce these complications, predictive modeling of radiation toxicity is helpful.Modeling based on radiomic features, in addition to being a simple, low-cost, and non-invasive method, is based on individual patient information, thus it can provide reliable prediction models with reproducibility and acceptable performance.
In this study, with the help of machine learning models based on the radiomic features of medical images and their combination with dosimetric and clinical parameters of patients, we showed that the radiomic features of MR images perform better than CT for the development of rectal radiation toxicity prediction models.Furthermore, when the radiomic features of images are combined with the clinical and dosimetric parameters of patients, the performance of predictive models improves.And finally, according to the models developed in this study, KNN and RF models can be used with acceptable performance to predict rectal radiation toxicity.

For
radiomic analysis in this work, CT and MR images were employed.The images were loaded into the open-source software program 3D slicer v.4.11 for radiomic analysis.Using the 3D-Slicer software, manually produced regions of interest (ROI) were established, encompassing the rectal wall but omitting the rectal lumen.ROIs were drawn on every slice and volumes of interest (VOIs) were made and then utilized for the feature extraction.All segmentations were validated by a skilled radiologist.The rectal wall's indicated ROIs are seen in Fig. 1.(a) (b).

Fig. 1
Fig. 1 Segmentation of ROI on the rectal wall at (a) CT and (b) MR images

Fig. 2
Fig. 2 Selected features using the LASSO model in (a) radiomic features of CT images, (b) radiomic features of MR images, (c) radiomic features of CT images along with clinical parameter, (d) radiomic features of MR images along with clinical parameter, (e) radiomic features of CT images along with dosimetric parameter, (f) radiomic

Fig. 3
Fig. 3 Heat map of (a) AUC, (b) Accuracy, (c) Sensitivity, and (d) Specificity for different combinations of models and feature sets

Fig. 4
Fig. 4 ROC curve of KNN models based on radiomic features of (a) CT images, (b) MR images, (c) CT images in combination with dosimetric and clinical features, and (d) MR images in combination with dosimetric features

Fig. 5
Fig. 5 Box plot of the (a) AUC, (b) Accuracy, (c) Sensitivity, and (d) Specificity for different feature sets

Table 2
Clinical parameters (Patient characteristics)