Machine Learning-Based Radiomics Model to Predict Benign and Malignant PI-RADS v2.1 Category 3 lesions : A Retrospective Multi-center Study

Machine Learning-Based Radiomics Model to Predict Benign and Malignant PI-RADS v2.1 Category 3 lesions : A Retrospective Multi-center Study | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Research Article Machine Learning-Based Radiomics Model to Predict Benign and Malignant PI-RADS v2.1 Category 3 lesions : A Retrospective Multi-center Study Pengfei Jin, Junkang Shen, Liqin Yang, Ji Zhang, Ao Shen, Jie Bao, and 1 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-2324823/v1 This work is licensed under a CC BY 4.0 License Status: Published Journal Publication published 29 Mar, 2023 Read the published version in BMC Medical Imaging → Version 1 posted 8 You are reading this latest preprint version Abstract Purpose: To develop machine learning-based prediction models derive from different MRI sequences for distinction between benign and malignant PI-RADS 3 lesions before intervention, and to cross-institution validate the generalization ability of the models. Methods: The pre-biopsy MRI datas of 463 patients diagnosed as PI-RADS 3 lesions were collected from 4 medical institutions. 2347 radiomics features were extracted from the VOI of T2WI, DWI and ADC maps. The ANOVA feature ranking method and support vector machine (SVM) classifier were used to construct 3 single-sequence models and 1 integrated model combined with the features of three sequences. All the models were established in the training set and independently verified in the internal test and external validation set. The area under the receiver operating characteristic curve (AUC) was used to quantify the predictive performance of each model. Hosmer-lemeshow test was used to evaluate the degree of fitting between prediction probability and pathological results. Non-inferiority test was used to check generalization performance of the integrated model. Results: T2WI-model with the mean AUC of 0.717 for predicting clinically significant prostate cancer (csPCa) (internal test AUC = 0.738 vs. external validation AUC = 0.695, P = 0.264) and 0.624 for predicting all cancer (internal test AUC = 0.678 vs. external validation AUC = 0.589, P = 0.547). DWI-model with the mean AUC of 0.658 for predicting csPCa (internal test AUC = 0.635 vs. external validation AUC = 0.681, P = 0.086) and 0.655 for predicting all cancer (internal test AUC = 0.712 vs. external validation AUC = 0.598, P = 0.437). ADC-model with the mean AUC of 0.746 for predicting csPCa (internal test AUC = 0.767 vs. external validation AUC = 0.724, P = 0.269) and 0.645 for predicting all cancer (internal test AUC = 0.650 vs. external validation AUC = 0.640, P = 0.848). Integrated model with the mean AUC of 0.803 for predicting csPCa (internal test AUC = 0.804 vs. external validation AUC = 0.801, P = 0.019) and 0.763 for predicting all cancer (internal test AUC = 0.801 vs. external validation AUC = 0.754, P = 0.047). Conclusions: The radiomics model based on mechine learning has the potential to be a non-invasive tool to distinguish cancerous, noncancerous and csPCa in PI-RADS 3 lesions and the generalization ability between different date set. Radiomics Clinically significant prostate cancer PI-RADS 3 Machine learning Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Figure 6 Introduction Prostate cancer (PCa) is a global public health problem that threatens human health and life, which causes great harm to the male genitourinary system[ 1 ]. According to statistics from the American Cancer Research Association and the National Cancer Institute in 2019, PCa has become one of the most common malignant tumors in the world, accounting for the second most common malignancy in men[ 2 ]. MRI imaging is a common and highly effective exam for the diagnosis of prostate diseases. PI-RADS v2.1 published by American College of Radiology in 2019, represents a standardized method for assessing and reporting prostate MRI, which categorizes prostate lesions into different classes to reflect their relative likelihood of clinically significant prostate cancer (csPCa)[ 3 ]. PI-RADS 3 lesions included benign lesions and malignant lesions with different invasiveness and due to the absence of a clear tendency diagnosis for PI-RADS 3 lesions, there is a great variability in the practice patterns of different institutions (from conservative treatment, imaging follow-up to targeted biopsy), expense and potential clinical results[ 4 ]. Studies on evaluating the possibility of csPCa in targeted biopsies of PI-RADS 3 lesions have reported that cancer diagnosis rates range from 5–30%, and most studies have suggested that the likelihood of eventual diagnosis of csPCa is relatively low[ 5 – 7 ]. Therefore, accurately judging the benign and malignant lesions is helpful to reduce the pain caused by unnecessary biopsies. Imaging monitoring without intervention for PI-RADS category 3 lesions will undoubtedly reduce unnecessary biopsies. However, this method may lead to omission or delay in the diagnosis of csPCa lesions, resulting in irreversible consequences for patients. There is still controversy over whether to intervene in this category of lesions[ 8 ], and the small but not insignificant proportion of lesions that represent csPCa, it is critical that a more detailed classification of the category 3 lesions will benefit patients from biopsies and more aggressive treatment. Radiomics can convert images to higher-dimensional data, extract a large number of phenotypic features, and evaluate the biological behavior of tumor noninvasively through machine learning (ML) algorithms. It has been widely used in the diagnosis, invasiveness evaluation and clinical decision-making of PCa[ 9 – 11 ]. The number of radiomics studies foucus on PI-RADS 3 lesions is limited only two single-center studies have previously assessed the role of radiomics characteristics to detect cancer in these “equivocal lesions”. However, there are doubts about the universality and wide applicability of radiomics models in the absence of multi-institution trials. Therefore, the purpose of this study was to construct a ML model combining T2WI, DWI and ADC radiomics features through a multi-center retrospective case-control study to validate its performance in differentiating PI-RADS 3 lesions from benign to malignant and in further risk stratification. Materials And Methods Study population and image acquisition This retrospective multi-agency study was approved by the ethics review committee of each participating institution and exempted from the need for informed consent of the patient. Four medical centers have signed data sharing agreements for data exchange (2021; Approval No. 262). All prostate MRI images from January 2018 to December 2019 were exported from each participating unit's PACS system. We summarized the data of each hospital, and there were a total of 2259 cases. 96 cases were excluded due to incomplete image sequence and lack of pathological data, then the remaining 2163 cases were divided into two parts and graded according to PI-RADS v2.1 criteria[3] by two radiologists with 3 years of experience in prostate MRI diagnosis, who were blind to pathological findings when reading. While interpreting the images, two radiologists recorded the location of each lesion using the anatomical fan map recommend by PI-RADS v2.1 to correspond to the lesion described by the pathological results. At an interval of two weeks after the first score, the procedure was repeated by two readers and reviewed by a senior radiologist proficient in MRI diagnosis of the urinary system. When there was any dispute over the interpretion, the three physicians discussed it until consensus was reached. Of the 2163 cases with final score results, 876 cases (40.5%) had PI-RADS scores of 1 and 2, 792 cases (36.6%) had PI-RADS scores of 4 and 5, and the remaining 495 cases (22.9%) were conferer with PI-RADS score 3. Then, all PI-RADS category 3 cases were selected for analysis, of which 32 were excluded based on the following criteria: (1) PI-RADS category 3 lesions coexisted with other types of lesions; (2) prior to MRI examination, they had received intervention such as biopsy, surgery or hormone therapy; (3) lack of any clinical characteristics of the patient or poor image quality. Finally, 463 eligible patients were recruited and MRI images of each patient showed only one lesion. All patients were divided into two groups according to the supplier of scanning equipment. The first group included institutions 1-3 with a total of 383 patients, which were examined with 3.0T superconducting MRI scanner (MAGNETOM Skyra, Germany) and equipped with 8-channel phased array body coils to collect signals. The second group consists of institution 4, with a total of 80 patients using a Dutch Philips Ingenia 3.0T MR scanner, the receiving coil was a 32-channel body phased array coil. The scanning sequences included T1WI, axial T2WI (no fat-saturated), sagittal T2WI, DWI (b=100, 800, 1500, 2000 s/mm 2 ) and / or dynamic contrast enhanced T1WI. According to the DWI images with high b value, the ADC value was calculated by extended single exponential fitting model. The case datas of the first group were randomly divided into training set (n = 268) and internal test set (n = 115) according to the proportion of 7:3. The second group of cases was used as an external validation set (n = 80) to evaluate the extensibility of the model. During DCE scanning, 15 to 20 slices were scanned once, the scanning time resolution was 5.8 s, 64 phases were scanned, and the scanning time was 7min. After the end of the third dynamic scanning phase, contrast agent gadolinium meglumine pentanoate was injected intravenously at the injection rate of 3 ml/s and the dose of 0.1 mmol/kg. MRI scan parameters are described in Table 1. Fig. 1 provides a flowchart that includes patient selection and case assignment. Targeted biopsy and histopathology MRI-TRUS fusion targeted sample was performed with Hitachi real-time ultrasonic multi-image fusion navigation system (RVS), and the machine model was HIVISIONNoblus/TopicPath. The suspicious lesions were sampled by MRI-TRUS fusion biopsy and systematic puncture under the guidance of TRUS within 4 weeks after the MRI examination. Before the fusion biopsy, the original data of prostate MRI in DICOM format were introduced into the main body of RVS ultrasound. MRI images were fused with TRUS images after general anesthesia, and anatomical markers such as urethra orifica, urethra, mullerian or ejaculatory duct cyst were matched with MRI sagittal images on the same section. T2WI, DWI, or DCE images with significant abnormal signals were selected to mark the target lesions in the cross-sectional MRI, while the same ultrasound sites were labeled (convex array scan), and then switched to sagittal images to further confirm the synchronization of MRI and ultrasound. After confirming favourable synchronization of MRI-TRUS images, the sagittal plane of prostate was taken by TRUS, and the target lesion marked with "+" was found. Under the guidance of puncture stent, the 18G disposable puncture gun was used to insert needle through perineum and the puncture gun was fired close to the target center. Then, the axial plane scan was converted to confirm that the needle track enters the target. 2-4 needles were punctured for each suspicious focus. After the targeted puncture, 12-needle puncture was conducted through perineum under the guidance of TRUS, and the pathological specimens were marked in detail according to each partition and fixed with 10% formaldehyde for pathological examination. The pathological results were evaluated by urological pathologist independently of the results of MRI, and the location and boundary of the lesions were recorded to ensure that they correspond to the suspicious lesions on MRI maps. The grade grouping and Gleason score of the lesions were determined according to the 2014 ISUP guidelines. csPCa was defined as ISUP Class 2 or higher (Gleason=3+4 or higher), and pathological results with GS=3+3 (ISUP class 1) were defined as clinically insignificant PCa (ciPCa)[12]. MRI Image preprocessing and focus segmentation Subsequent evaluation and analysis were performed based on axial T2WI, DWI (b=2000 s/mm 2 ) and ADC sequences in our study. The target images were spatially matched to ensure that DWI and ADC have the same resolution, field of view (FOV) and orientation compared with T2WI by using Elastix software package (v.4.10, http://elastix.isi.uu.nl/index.php)[13]. Before the focus segmentation, the images were standardized to improve the texture recognition rate. The μ±3σ method was used to normalize the image, remove the gray signal more than 3 σ, and quantify the gray level with 64 levels. Finally, the voxel size of 1mm × 1mm × 1mm was used for equal-view resampling. The resampling operation was done by the "Radiomics" package of R software. Lesion segmentation was performed jointly by two radiologist involved in image evaluation using ITK-SNAP 3.8.0 software (http://www.itksnap.org/). The two handlers drew the region of interest (ROI) layer by layer on T2WI sequence to get the volume of interest (VOI) of the tumor, then copy it to DWI and ADC images to ensure the consistency of VOI sketches in different sequences. After preprocessing, visually verified was performed by a professor with experience in prostate MRI diagnosis (more than 10 years) to ensure that the location and extent of the lesions shown on MRI strictly matched the corresponding pathological description. Feature extraction and stability evaluation The open source radiomics software FeAture Explorer (FAE v0.4.0) developed by Song et al was used to extract features from the VOI of each image sequence[14]. According to the 8 texture analysis methods provided by the software, a total of 2347 image features were extracted from ROI files: (1) 46 first-order gray statistics; (2) 38 shape-based features; (3) 70 Gray Level Co-occurrence Matrices (GLCM); (4) 20 Gray Level Run Length Matrices (GLRLM); (5) 42 Gray Level Size Zone Matrices (GLSZM); (6) 36 Gray Level Dependence Matrices (GLDM); and (7) 17 Neighborhood Gray Tone Difference Matrices (NGTDM). (8) the original images were transformed by Wavelet Transform, and 2078 wavelet features are extracted in three spatial directions. The repeatability of intra- and inter-observer of lesion segmentation was based on the repeatability of feature extraction. 30 patients were randomly selected and the clinical data were blinded. The two doctors performed VOI segmentation and feature extraction again. The intra- and inter-observer repeatability of feature extraction was evaluated by intergroup correlation coefficient (ICC). If the intra-group and inter-group correlation coefficient is greater than 0.75, it is considered that the ROI drawing has acceptable stability. Feature selection and classifier modeling In this study, we focus on two results: (1) distinguish any cancer diagnosis from benign diagnosis, (2) and further predict csPCa occurrence in all cases. In order to solve the problem of sample imbalance in the characteristic matrix, this study uses the synthetic minority oversampling technique (SMOTE) to oversample the unbalanced positive and negative samples in the training set to balance the data set[15]. The number of image features was much larger than the number of samples, which may increase the risk of overfitting. This risk was reduced by feature selection to reduce the number of features. In present study, Z-Score Normalization was first used to normalize the feature matrix, each feature vector was subtracted from the mean value and divided by the standard deviation to eliminate the order of magnitude otherness between different features. The image features with variance of 0 were eliminated, and then the data dimension was reduced to remove the redundant features with average Spearman absolute correlation coefficient ≥ 0.9. After eliminated redundant features, the analysis of variance (ANOVA) algorithm was used to sort the features, and only the top 20 features were retained. These features with increments from 1 to 20 were then input into the support vector machine (SVM) classifier. For different sequences, T2WI, DWI and ADC feature matrices were modeled respectively (called T2WI-model, DWI-model and ADC-model), and then the features of the three sequences were combined for modeling analysis (call integrated model). While established models to identify csPCa, the features of the first group were re-integrated, and the features of benign lesions and ciPCa were divided into the same group, and also divided into training set (n=268) and internal test set (n=115) according to the proportion of 7:3, and the generalization ability of the model was verified on the external valitation set. All the experiments above were run in FeAtureExplorer. Statistical analysis Demographic datas were compared by chi-square test and independent t-test. According to whether it conformed to the normal distribution, the quantitative data were expressed as average (± standard deviation) or median (quartile range). Prediction models were inspected on the internal test and external validation sets. The receiver operating characteristic (ROC) curve was analyzed and the area under the ROC curve (AUC) was quantified to evaluate their performance in distinguishing cancer from benign lesions. Hosmer-lemeshow test was used to evaluate the degree of fitting between the predicted results of the integrated model and the histopathological results, and drawn the calibration diagram to visually display the results. In order to evaluate the generalization ability of the model, the non-inferior test was used to check whether the AUC of the external validation set is not lower than that of in the internal test set. R software (version 4.1.0, www. Rproject. org) was uesd for non-inferiority testing, the predefined acceptable threshold value was set to 0.1. Through the non-inferiority test of each model, the P-value was obtained, when P < 0.05, it indicates that the model has good versatility. Results Clinical characteristics included only age and prostate specific antigen (PSA). The mean age of patients was 62.6±8.2 years, and PSA was 9.5±6.9 ng/mL. Of the 463 PI-RADS v2.1 category 3 lesions, 311 (67.2%) were benign and 152 (32.8%) were PCa lesions, of which 11.2%(52/463) were ciPCa (ISUP grade 1), 21.6% (100/463) were csPCa (47 ISUP level 2, 20 ISUP Level 3, 23 ISUP Level 4, 10 ISUP Level 5). There was no difference in the distribution of PCa and csPCa between different institutions ( P =0.502, 0.173). From the 463 PI-RADS category 3 lesion, there were 216 peripheral zone lesions (46.7%) with 79 PCa (48 csPCa and 31 ciPCa) and 247 transition zone lesions (53.3%) with 73 PCa (52 csPCa and 21ciPCa). The patient's demographic and clinical datas were shown in Table 2. In the intra- and inter-observer consistency test, the intra-observer ICCs range was 0.77 to 0.90, and the inter-observer ICCs range was 0.80 to 0.87, indicated that the repeatability of feature extraction was fine. Spearman correlation test results of the top 20 features screened by ANOVA were represented by feature heat map (Fig. 2). While constructed the integrated model, 6 and 5 features were screened to distinguish benign from malignant lesions and to further identify csPCa in cancer lesions. The name of the features and the corresponding coefficient are shown in fig 3. The accuracy of T2WI-model in identifying csPCa of PI-RADS 3 lesions was 0.774 and 0.763 in internal test and external validation set, respectively, and the mean AUC value was 0.717 (internal test AUC = 0.738, external valitation AUC = 0.695, P = 0.547). The accuracy of the model in distinguishing benign and malignant PI-RADS 3 lesions in internal test and external validation set was 0.644 and 0.650, respectively, with mean AUC of 0.624. The accuracy of DWI-model in identifying csPCa of PI-RADS 3 lesions was 0.730 and 0.813 in internal test and external validation set, respectively, and the mean AUC value was 0.658 (internal test AUC = 0.635, external validation AUC = 0.681, P = 0.547). The accuracy of the model in distinguishing benign and malignant PI-RADS 3 lesions in internal test and external validation set was 0.730 and 0.638, respectively, with mean AUC of 0.655. The accuracy of ADC-model in identifying csPCa of PI-RADS 3 lesions was 0.739 and 0.775 in internal test and external validation set, respectively, and the mean AUC value was 0.746 (internal test AUC = 0.767, external validation AUC = 0.724, P = 0.547). The accuracy of the model in distinguishing benign and malignant PI-RADS 3 lesions in internal test and external validation set was 0.565 and 0.613, respectively, with mean AUC value of 0.645. The integrated model based on three single-sequence radiomics features, and its accuracy in identifying csPCa was 0.748 in internal test set and 0.863 in external validation set. The mean AUC value was 0.803 (internal test AUC = 0.804, external validation AUC = 0.801, P = 0.019). The accuracy of the model in distinguishing benign and malignant PI-RADS 3 lesions in internal test and external validation set was 0.748 and 0.762, respectively, with mean AUC of 0.763. The results of Hosmer-Lemeshow test showed that the prediction results of the integrated model for all PCa and csPCa in the internal test and the external validation set had a high coincidence rate with the observed risks (P = 0.073 vs. 0.082 for PCa; P = 0.224 vs. 0.647 for csPCa, respectively). The results of each model for distinguishing benign and malignant PI-RADS 3 diseases are shown in Table 3, and the corresponding ROC curves are shown in Fig. 4. The effectiveness of each model in identifying csPCa is compared in Table 4, and the corresponding ROC curves are shown in fig. 5. The pathological calibration scatter plots of the prediction results of the integrated model are shown in Fig. 6. Discussion This study shows that ML models based on T2WI, DWI and ADC radiomics features can achieve upper-moderate accuracy when predicting any cancer and csPCa in PI-RADS v2.1 3 lesions, and the performance of integrated model is better than that of all single-sequence models, which indicates that only based on the simplex radiomics feature may be limited in distinguishing significant tumors from benign or inert lesions, and the combination of multiple features is well complementary. However, it is worth noting that the performance of all models in predicting csPCa is better than that of models in predicting all cancers. Therefore, our results also show that the heterogeneity of csPCa is more obvious than that of ciPCa, and it is easier to be recognized in ML progress. Several additional indicators have been introduced to predict the need for biopsy in patients with PI-RADS 3, including lesion size, prostate volume, ADC, PSA and PSA density, but the published results do not fully prove the relationship between these indicators and the risk of csPCa appearance[ 16 – 19 ]. For example, quantitative ADC values can help detect carcinoma while avoiding biopsies that are negative[ 20 ]. Zhang et al showed that age, PSA density, lesion zone and ADC value were predictors of csPCa and PCa[ 21 ]. However, another study showed that the difference of median ADC values in PI-RADS 3 lesions was not statistically significant[ 22 ]. In recent years, radiomics studies have mainly focused on tumor detection, prediction of PI-RADS score and Gleason grade, evaluation of tumor extra-capsular extension and therapeutic response, which have shown similar performance as PI-RADS[ 23 – 24 ]. However, there are few studies use radiomics to assisted diagnose PI-RADS 3 lesions, and lack multi-center studies to validate the generalization ability of the model. Our results show that the single-sequence model is less efficient in both internal test and external validation set, with the lowest mean AUC for T2WI radiomics features, which is similar to the results of Lim et al[ 25 ]. They constructed a model based on XGBoost algorithm to predict any cancer or csPCa in PI-RADS 3 lesions, and AUC performed by T2WI features for all types of tumor was 0.608 and 0.547 for csPCa, lower than 0.642 and 0.684 of ADC features. Hectors et al. reconfirmed that model with T2WI radiomics features had a low ability to diagnose csPCa (AUC = 0.76)[ 26 ]. However, ADC and DWI characteristics were not included as controls in their study. Our results are lower than those of Hou et al[ 27 ], who extracted features from T2WI, DWI and ADC images, constructed a one-step ML model and a regression analysis model integrated radiomics score, and improved the risk stratification method for identifying csPCa in PI-RADS 3 lesions with AUC reached 0.74 to 0.89. There are several design differences between this study and previous studies, which may explain the conflicts in results with Hou and Hectors. In contrast to these studies, our study used MRI datas from two vendors in four medical units. Different MRI scanners are equipped with different software and hardware, and these differences mean that scanners may not obtain images with the same intensity distribution. Although we performed resampling and registration operations before model construction, there is still the possibility of affecting the performance and generalization ability of the model. For this reason, an independent external validation set was set up to evaluate the model's performance, and this group of cases was provided by a different supplier from the testing set[ 28 – 30 ]. The models constructed by Hou et al and Hectors et al were trained and tested only in their respective institutions, which limited extensibility. For example, quantitative values of DWI and ADC may be affected by variabilities between different scanners, imaging parameters, and patients, which caused the repeatability controversial. The lower accuracy of our study may be due to the fact that datas from multiple centers were integrated together and the number of PCa contributed by each participating unit was different, leading to differences in the distribution of cases. In order to ensure consistency between the combined data set and the distribution of cases in a single center, Lim et al. conducted a subgroup analysis of larger disease-causing institutions, but was unable to confirm this conjecture. Our study used non-inferiority test to evaluate the model's generalization ability, which was not available in other studies. Although we failed to prove that the AUC of all single-sequence radiomics featuers in the external validation set was not lower than that in the internal test set ( P > 0.05). However, the diagnostic accuracy and sensitivity of the integrated model in external validation set are higher than that of the internal test set, and the AUC in external validation set was not inferior to the AUC in internal test set ( P < 0.05), indicating that the integrated model has certain generalization ability in different date sets. In addition, Ji et al constructed a comprehensive model combinie age, PSA and radiomics features, suggested that combin clinical features can improve the generalization ability of radiomics model[ 31 ]. Different reference standards may also be one of the reasons for the different results. In Hou et al. 's study, a subset of included lesions lacked pathological diagnosis, and the clinical significance of tumor foci was only inferred based on follow-up imaging results and/or PSA changes after empirical treatment. This limits the reliability of the model's results for predicting a subset of clinically significant cancers, some of which were misclassified when they could have been monitored closely[ 32 ]. For the single-sequence model, the features based on DWI/ADC sequence are better than that based on T2WI in distinguishing between benign and malignant lesions. This is consistent with the research of Hou et al. In anther similar study, the most important feature for detecting tumor in PI-RADS 3 lesions was based on ADC maps[ 33 ]. The changes of diffusion of water molecules in tissues were monitored by DWI images, reflecting the changes of cell volume and number in epithelium, stroma and luminal space[ 34 ]. PCa is high cellular tissue, which restricts the diffusion to some extent due to the blocking of the random movement of water molecules in the tumor. The degree of diffusion limitation is positively correlated with the tumor grade, invasiveness and stage[ 35 ]. ML can quantify subtle changes in the diffusion motion of water molecules in the DWI/ADC diagram, which makes DWI perform better than other sequences to evaluate the PCa. There are several limitations in this study. First, this is a retrospective case-control study with a relatively small sample size, especially for a small number of csPCa with uneven distribution between groups, may be at risk of over-fitting when training models, which limits the evaluation of the accuracy in predicting malignant tumor; second, the sketch of ROI was done by radiologists by hand, which is time-consuming and affected by inter-observer variability. Automatic ROI segmentation algorithm can be introduced to improve the speed in the future. Third, the location of the lesion, such as peripheral and transitional zones, or poorly defined areas, was not taken into account. Due to the differences between peripheral zones and transition zones, modeling for each region may affect model performance. Finally, it was not discussed whether the clinical factors combined with radiomics features can provide additional diagnostic value for PI-RADS 3 lesions. Conclusion The radiomics feature-based mechine learning model achieved an encouraging performance in differentiating PI-RADS 3 lesions from benign to malignant and distinguishing significant or indolent tumors, which has certain application value to assist clinical decision making, and provides a new direction for the management of patients with controversial MRI diagnosis and helps to reduce unnecessary biopsies while improving the detection rate of csPCa. Abbreviations PCa prostate cancer; PI-RADS prostate imaging report and data system; csPCa clinically significant prostate cancer; ML machine learning; ROI region of interest; VOI volume of interest; ROC receiver operating characteristic; AUC area under curve Declarations Acknowledgements The authors thank all those who helped us during the writing of this research. We also thank the Department of Urology and Pathology of the hospitals for their valuable help and feedback. Author contributions Pengfei Jin: manuscript drafting and revision, study concept and design, collection, assembly, interpretation of the data, and figure drawing. Liqin Yang:manuscript drafting and figure drawing. Junkang Shen, Ao Shen and Ji Zhang: data collection and interpretation of the data. Jie Bao and Ximing Wang: Manuscript writing and final approval of the manuscript. All authors read and approved the final manuscript. Fundings This study was supported by the Special Program for Diagnosis and Treatment Technology of Clinical Key Diseases in Suzhou (LCZX202001), Gusu health talent project of Suzhou (GSWS2020003), Suzhou Key Laboratory of health information technology (SZS201818). Availability of data and materials The datasets generated during this study are available from the corresponding author upon reasonable request. Ethical approval and Consent to Participate This study was under ethics approval of the First Afliated Hospital of Soochow University (Approval No. 262; 2021). All procedures performed in studies involving human participants were in accordance with the ethical standards of the institutional and/or national research committee and with the 1964 Helsinki declaration and its later amendments or comparable ethical standards. Informed consent was obtained from all individual participants included in the study. Consent for publication Not applicable. Competing interests All authors report no conficts of interest. References Teoh JYC, Hirai HW, Ho JMW, et al. Global incidence of prostate cancer in developing and developed countries with changing age structures. PLoS One. 2019 Oct 24;14(10):e0221775. doi: 10.1371/journal.pone.0221775. Culp MB, Soerjomataram I, Efstathiou JA. Recent Global Patterns in Prostate Cancer Incidence and Mortality Rates. Eur Urol. 2020 Jan;77(1):38-52. doi: 10.1016/j.eururo.2019.08.005. Turkbey B, Rosenkrantz AB, Haider MA, et al.Prostate Imaging Reporting and Data System Version 2.1: 2019 Update of Prostate Imaging Reporting and Data System Version 2. Eur Urol,2019,76(3):340-351. doi: 10.1016/j.eururo.2019.02.033. Maggi M, Panebianco V, Mosca A, et al. Prostate Imaging Reporting and Data System 3 Category Cases at Multiparametric Magnetic Resonance for Prostate Cancer: A Systematic Review and Meta-analysis. Eur Urol Focus. 2020 May 15;6(3):463-478. doi: 10.1016/j.euf.2019.06.014. Liddell H, Jyoti R, Haxhimolla HZ. mp-MRI Prostate Characterised PIRADS 3 Lesions are Associated with a Low Risk of Clinically Significant Prostate Cancer - A Retrospective Review of 92 Biopsied PIRADS 3 Lesions. Curr Urol. 2015 Jul;8(2):96-100. doi: 10.1159/000365697. Schoots IG. MRI in early prostate cancer detection: how to manage indeterminate or equivocal PI-RADS 3 lesions? Transl Androl Urol. 2018 Feb;7(1):70-82. doi: 10.21037/tau.2017.12.31. Otti VC, Miller C, Powell RJ, et al. The diagnostic accuracy of multiparametric magnetic resonance imaging before biopsy in the detection of prostate cancer. BJU Int. 2019 Jan;123(1):82-90. doi: 10.1111/bju.14420. Lim CS, Abreu-Gomez J, Leblond MA, et al. When to biopsy Prostate Imaging and Data Reporting System version 2 (PI-RADSv2) assessment category 3 lesions? Use of clinical and imaging variables to predict cancer diagnosis at targeted biopsy. Can Urol Assoc J. 2021 Apr;15(4):115-121. doi: 10.5489/cuaj.6781. Khalvati F, Zhang J, Chung AG, et al. MPCaD: a multi-scale radiomics-driven framework for automated prostate cancer localization and detection. BMC Med Imaging. 2018,18(1):16. doi: 10.1186/s12880-018-0258-4. Wang J, Wu CJ, Bao ML, et al. Machine learning-based analysis of MR radiomics can help to improve the diagnostic performance of PI-RADS v2 in clinically relevant prostate cancer. Eur Radiol. 2017 Oct;27(10):4082-4090. doi: 10.1007/s00330-017-4800-5. Min X, Li M, Dong D, et al. Multi-parametric MRI-based radiomics signature for discriminating between clinically significant and insignificant prostate cancer: Cross-validation of a machine learning method. Eur J Radiol. 2019 Jun;115:16-21. doi: 10.1016/j.ejrad.2019.03.010. Epstein JI, Egevad L, Amin MB, et al. The 2014 International Society of Urological Pathology (ISUP) Consensus Conference on Gleason Grading of Prostatic Carcinoma: Definition of Grading Patterns and Proposal for a New Grading System. Am J Surg Pathol. 2016 Feb;40(2):244-52. doi: 10.1097/PAS.0000000000000530. Klein S, Staring M, Murphy K, et al. elastix: a toolbox for intensity-based medical image registration. IEEE Trans Med Imaging. 2010 Jan;29(1):196-205. doi: 10.1109/TMI.2009.2035616. Song Y, Zhang J, Zhang YD, et al. FeAture Explorer (FAE): A tool for developing and comparing radiomics models. PLoS One. 2020 Aug 17;15(8):e0237587. doi: 10.1371/journal.pone.0237587. Seo JH, Kim YH. Machine-Learning Approach to Optimize SMOTE Ratio in Class Imbalance Dataset for Intrusion Detection. Comput Intell Neurosci. 2018 Nov 1;2018:9704672. doi: 10.1155/2018/9704672. Felker ER, Raman SS, Margolis DJ, et al. Risk Stratification Among Men With Prostate Imaging Reporting and Data System version 2 Category 3 Transition Zone Lesions: Is Biopsy Always Necessary? AJR Am J Roentgenol. 2017 Dec;209(6):1272-1277. doi: 10.2214/AJR.17.18008. Washino S, Okochi T, Saito K, et al. Combination of prostate imaging reporting and data system (PI-RADS) score and prostate-specific antigen (PSA) density predicts biopsy outcome in prostate biopsy naïve patients. BJU Int. 2017 Feb;119(2):225-233. doi: 10.1111/bju.13465. Hermie I, Van Besien J, De Visschere P, et al. Which clinical and radiological characteristics can predict clinically significant prostate cancer in PI-RADS 3 lesions? A retrospective study in a high-volume academic center. Eur J Radiol. 2019 May;114:92-98. doi: 10.1016/j.ejrad.2019.02.031. Yang S, Zhao W, Tan S, et al. Combining clinical and MRI data to manage PI-RADS 3 lesions and reduce excessive biopsy. Transl Androl Urol. 2020 Jun;9(3):1252-1261. doi: 10.21037/tau-19-755. Girometti R, Giannarini G, Panebianco V, et al. Comparison of different thresholds of PSA density for risk stratification of PI-RADSv2.1 categories on prostate MRI. Br J Radiol. 2021 Nov 11:20210886. doi: 10.1259/bjr.20210886. Zhang Y, Zeng N, Zhang FB, et al. Performing Precise Biopsy in Naive Patients With Equivocal PI-RADS, Version 2, Score 3, Lesions: An MRI-based Nomogram to Avoid Unnecessary Surgical Intervention. Clin Genitourin Cancer. 2020 Oct;18(5):367-377. doi: 10.1016/j.clgc.2019.11.011. Hermie I, Van Besien J, De Visschere P, et al. Which clinical and radiological characteristics can predict clinically significant prostate cancer in PI-RADS 3 lesions? A retrospective study in a high-volume academic center. Eur J Radiol. 2019 May;114:92-98. doi: 10.1016/j.ejrad.2019.02.031. Chaddad A, Niazi T, Probst S, et al. Predicting Gleason Score of Prostate Cancer Patients Using Radiomic Analysis. Front Oncol. 2018 Dec 18;8:630. doi: 10.3389/fonc.2018.00630. Gong L, Xu M, Fang M, Zou J, et al. Noninvasive Prediction of High-Grade Prostate Cancer via Biparametric MRI Radiomics. J Magn Reson Imaging. 2020 Oct;52(4):1102-1109. doi: 10.1002/jmri.27132. Lim CS, Abreu-Gomez J, Thornhill R, et al. Utility of machine learning of apparent diffusion coefficient (ADC) and T2-weighted (T2W) radiomic features in PI-RADS version 2.1 category 3 lesions to predict prostate cancer diagnosis. Abdom Radiol (NY). 2021 Dec;46(12):5647-5658. doi: 10.1007/s00261-021-03235-0. Hectors SJ, Chen C, Chen J, et al. Magnetic Resonance Imaging Radiomics-Based Machine Learning Prediction of Clinically Significant Prostate Cancer in Equivocal PI-RADS 3 Lesions. J Magn Reson Imaging. 2021 Nov;54(5):1466-1473. doi: 10.1002/jmri.27692. Hou Y, Bao ML, Wu CJ, et al. A radiomics machine learning-based redefining score robustly identifies clinically significant prostate cancer in equivocal PI-RADS score 3 lesions. Abdom Radiol (NY). 2020 Dec;45(12):4223-4234. doi: 10.1007/s00261-020-02678-1. Litjens GJ, Hambrock T, Hulsbergen-van de Kaa C, et al. Interpatient variation in normal peripheral zone apparent diffusion coefficient: effect on the prediction of prostate cancer aggressiveness. Radiology. 2012 Oct;265(1):260-6. doi: 10.1148/radiol.12112374. Barrett T, Lawrence EM, Priest AN, et al. Repeatability of diffusion-weighted MRI of the prostate using whole lesion ADC values, skew and histogram analysis. Eur J Radiol. 2019 Jan;110:22-29. doi: 10.1016/j.ejrad.2018.11.014. Schmeel FC. Variability in quantitative diffusion-weighted MR imaging (DWI) across different scanners and imaging sites: is there a potential consensus that can help reducing the limits of expected bias? Eur Radiol. 2019 May;29(5):2243-2245. doi: 10.1007/s00330-018-5866-4. Ji X, Zhang J, Shi W, et al. Bi-parametric magnetic resonance imaging based radiomics for the identification of benign and malignant prostate lesions: cross-vendor validation. Phys Eng Sci Med. 2021 Sep;44(3):745-754. doi: 10.1007/s13246-021-01022-1. Balakrishnan AS, Cowan JE, et al. Evaluating the Safety of Active Surveillance: Outcomes of Deferred Radical Prostatectomy after an Initial Period of Surveillance. J Urol. 2019 Sep;202(3):506-510. doi: 10.1097/JU.0000000000000247. Zhang KS, Schelb P, Kohl S, et al. Improvement of PI-RADS-dependent prostate cancer classification by quantitative image assessment using radiomics or mean ADC. Magn Reson Imaging. 2021 Oct;82:9-17. doi: 10.1016/j.mri.2021.06.013. Surov A, Meyer HJ, Wienke A. Correlations between Apparent Diffusion Coefficient and Gleason Score in Prostate Cancer: A Systematic Review. Eur Urol Oncol. 2020 Aug;3(4):489-497. doi: 10.1016/j.euo.2018.12.006. Brancato V, Aiello M, Basso L, et al. Evaluation of a multiparametric MRI radiomic-based approach for stratification of equivocal PI-RADS 3 and upgraded PI-RADS 4 prostatic lesions. Sci Rep. 2021 Jan 12;11(1):643. doi: 10.1038/s41598-020-80749-5. Tables Table 1 MRI protocols for both vender MRI vendor Sequence Siemens Skyra 3.0T MR scanner (Germany) Philips Ingenia 3.0 T MR scanner (Netherlands) T1WI AxialT2WI SagittlT2WI DWI T1WI AxialT2WI SagittlT2WI DWI TR(ms) 680.0 6980.0 3900.0 5000.0 556.0 3000.0 4978.0 6000.0 TE(ms) 13.00 104.00 89.00 72.00 8.00 100.00 100.00 77.00 Slice thickness(mm) 5.0 3.0 3.0 3.0 5.0 3.0 1.5 3.0 Slice gap(mm) 0.50 0.00 0.45 0.00 0.00 0.00 0.15 0.00 Matrix 384×384 384×384 384×384 130×130 276×406 240×161 276×238 124×121 FOV(mm×mm) 380×380 200×200 200×200 288×288 249×415 220×220 240×180 220×220 NSA 1 2 3 2 1 3 2 2 TR repetition time; TE echo time; NSA number of signal averaged; T1WI T1 weighted imaging; T2WI T2 weighted imaging; DWI Diffusion Weighted Imaging Table 2a Patient profiles of subgroup for distinguish any cancer diagnosis from benign lesions Training set Internal test set P-value External validation set P-value* Ages 66.5±10.4 69.8±7.6 0.108 70.2±12.4 0.257 PSA 13.3±11.8 14.7±9.3 0.429 11.2±7.3 0.315 Lesion type Benign Malignant (ciPCa+csPCa) Total 180 88 (30+58) 268 77 38 (13+25) 115 - 54 26 (9+17) 80 - PSA Prostate Specific Antigen; csPCa clinically significant prostate cancer; ciPCa non-clinically significant prostate cancer *The P values are derived from the comparison between training set, internal test set and external validation set Table 2b Patient profiles of each subgroup for predicting clinically significant prostate cancer Training set Internal test set P-value External validation set P-value* Ages 68.1±13.5 66.7±5.9 0.612 69.3±4.4 0.316 PSA 13.9±12.4 12.6±10.4 0.167 13.4±5.1 0.121 Lesion type non-csPCa (Benign+ciPCa) csPCa Total 210 (180+30) 58 268 90 (77+13) 25 115 - 63 (54+9) 17 80 - PSA Prostate Specific Antigen; csPCa clinically significant prostate cancer; ciPCa non-clinically significant prostate cancer *The P values are derived from the comparison between training set, internal test set and external validation set Table 3 The performance of each model for predicting any tumors in PI-RADS 3 lesions Modality Training set Internal test set External validation set Mean AUC* P value AUC ACC SEN SPE AUC ACC SEN SPE AUC ACC SEN SPE T2WI-model 0.811 0.784 0.614 0.867 0.678 0.644 0.842 0.545 0.589 0.650 0.500 0.722 0.624 0.547 DWI-model 0.717 0.735 0.557 0.822 0.712 0.730 0.684 0.753 0.598 0.638 0.615 0.648 0.655 0.437 ADC-model 0.840 0.780 0.773 0.783 0.650 0.565 0.921 0.390 0.640 0.613 0.654 0.593 0.645 0.848 Integrated-model 0.855 0.746 0.921 0.661 0.801 0.748 0.763 0.740 0.754 0.762 0.846 0.722 0.763 0.047 T2WI T2 weighted imaging; DWI Diffusion Weighted Imaging; ADC Apparent Diffusion Coefficient; AUC area under the receiver operating characteristic curve; ACC accuracy; SEN sensitivity; SPE specificity *Mean AUC = [AUC(Internal test set) + AUC(External validation set)]/2 The P-values from the non-inferiority tests Table 4 The performance of each model for predicting csPCa in all PI-RADS 3 lesions modality Training set Inner test set External validation set Mean AUC* P value AUC ACC SEN SPE AUC ACC SEN SPE AUC ACC SEN SPE T2WI-model 0.740 0.668 0.793 0.633 0.738 0.774 0.680 0.800 0.695 0.763 0.708 0.778 0.717 0.264 DWI-model 0.798 0.802 0.690 0.833 0.635 0.730 0.440 0.811 0.681 0.813 0.471 0.905 0.658 0.086 ADC-model 0.805 0.784 0.655 0.819 0.767 0.739 0.760 0.733 0.724 0.775 0.588 0.825 0.746 0.269 Integrated-model 0.854 0.828 0.741 0.852 0.804 0.748 0.800 0.733 0.801 0.863 0.921 0.647 0.803 0.019 csPCa clinically significant prostate cancer; T2WI T1 weighted imaging; DWI Diffusion Weighted Imaging; ADC Apparent Diffusion Coefficient; AUC area under the receiver operating characteristic curve; ACC accuracy; SEN sensitivity; SPE specificity *Mean AUC = [AUC(Internal test set) + AUC(External validation set)]/2 The P-values from the non-inferiority tests Additional Declarations No competing interests reported. Cite Share Download PDF Status: Published Journal Publication published 29 Mar, 2023 Read the published version in BMC Medical Imaging → Version 1 posted Editorial decision: Major revision 12 Jan, 2023 Reviews received at journal 22 Dec, 2022 Reviewers agreed at journal 19 Dec, 2022 Reviewers invited by journal 19 Dec, 2022 Editor assigned by journal 15 Dec, 2022 Editor invited by journal 14 Dec, 2022 Submission checks completed at journal 14 Dec, 2022 First submitted to journal 29 Nov, 2022 You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-2324823","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":159957910,"identity":"fb3cf97e-77d4-4fbc-99a4-2ede9ad83c08","order_by":0,"name":"Pengfei Jin","email":"","orcid":"","institution":"Department of Radiology, The Cancer Hospital of the University of Chinese Academy of Science (Zhejiang Cancer Hospital), Institute of Basic Medicine and Cancer (IBMC), Chinese Academy of Science","correspondingAuthor":false,"submittingAuthor":false,"prefix":"","firstName":"Pengfei","middleName":"","lastName":"Jin","suffix":""},{"id":159957911,"identity":"5655521b-d817-4311-91ab-a83b8b1a5443","order_by":1,"name":"Junkang Shen","email":"","orcid":"","institution":"Department of Radiology, The Second Affiliated Hospital of Soochow University","correspondingAuthor":false,"submittingAuthor":false,"prefix":"","firstName":"Junkang","middleName":"","lastName":"Shen","suffix":""},{"id":159957916,"identity":"7b590367-3f72-4c4a-aad1-4570fd2aa990","order_by":2,"name":"Liqin Yang","email":"","orcid":"","institution":"Department of Radiology, The First Affiliated Hospital of SooChow University","correspondingAuthor":false,"submittingAuthor":false,"prefix":"","firstName":"Liqin","middleName":"","lastName":"Yang","suffix":""},{"id":159957920,"identity":"cfdbb860-3fc0-456a-b2c0-89fed8e71024","order_by":3,"name":"Ji Zhang","email":"","orcid":"","institution":"Department of Radiology, Taizhou people’s Hospital of Jiangsu Province","correspondingAuthor":false,"submittingAuthor":false,"prefix":"","firstName":"Ji","middleName":"","lastName":"Zhang","suffix":""},{"id":159957923,"identity":"074da6f8-628b-4102-b3c1-d1ab5f51eb97","order_by":4,"name":"Ao Shen","email":"","orcid":"","institution":"Suzhou Institute of Biomedical Engineering and Technology, Chinese Academy of Sciences","correspondingAuthor":false,"submittingAuthor":false,"prefix":"","firstName":"Ao","middleName":"","lastName":"Shen","suffix":""},{"id":159957926,"identity":"792da62f-0a97-4561-adfa-fd1a39dc855f","order_by":5,"name":"Jie Bao","email":"","orcid":"","institution":"Department of Radiology, The First Affiliated Hospital of SooChow University","correspondingAuthor":false,"submittingAuthor":false,"prefix":"","firstName":"Jie","middleName":"","lastName":"Bao","suffix":""},{"id":159957927,"identity":"5361002b-63eb-4c99-a8a9-f70166b27670","order_by":6,"name":"Ximing Wang","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAA+0lEQVRIie2RsWrDMBCGJQTJckarRfIQKgEHQ6hfxVDIZErHjHINWrMq0KUvkVkmQxa3Xg1ekjkeFLp2qOrSMXLGQPQtB8d993MSQh7PzcIRIITzg1ktgFJxtUKKB1Utp0zpq6PGchLI3YKLdGD//mN3gpftdE7f5QSqGjjS2Jwzh1I9L2PFW4jVsYjfVi3MiSBss72sRDqLuLEKb8q86SrrCj0igUupu4inf4oIA/kJXKcDSpPNDn1Knb+yQOphJWm6CKs+Bf8+8hMwVRbOW9g6m33Bd5vwen+0X/mYUFqU5uxQLKOwL2H638DCOW8hpi9UDw16PB7PvfIDwgBcJ1kL0ZEAAAAASUVORK5CYII=","orcid":"","institution":"Department of Radiology, The First Affiliated Hospital of SooChow University","correspondingAuthor":true,"submittingAuthor":false,"prefix":"","firstName":"Ximing","middleName":"","lastName":"Wang","suffix":""}],"badges":[],"createdAt":"2022-11-29 10:44:17","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-2324823/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-2324823/v1","draftVersion":[],"editorialEvents":[{"content":"https://doi.org/10.1186/s12880-023-01002-9","type":"published","date":"2023-03-29T20:14:50+00:00"}],"editorialNote":"","failedWorkflow":false,"files":[{"id":30436000,"identity":"62f4dd33-19ff-4a42-8595-fbe47a0a6eca","added_by":"auto","created_at":"2022-12-16 16:11:34","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":311241,"visible":true,"origin":"","legend":"\u003cp\u003eFlow diagram on methods of this study\u003c/p\u003e","description":"","filename":"Fig.1.png","url":"https://assets-eu.researchsquare.com/files/rs-2324823/v1/71a300a74649bc4f584bec03.png"},{"id":30434906,"identity":"1c07c0f0-0e83-4ea4-9a8c-fdc33f81b6b4","added_by":"auto","created_at":"2022-12-16 16:03:34","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":18866,"visible":true,"origin":"","legend":"\u003cp\u003eDetermine the number of features used to construct the prediction model for differential diagnosisbetween benign and malignant lesions (a) and further identification csPCa in cancer lesions (b)\u003c/p\u003e","description":"","filename":"Fig.2ab.png","url":"https://assets-eu.researchsquare.com/files/rs-2324823/v1/0a0cb53bc3dbb32c075bd8a7.png"},{"id":30434905,"identity":"1a292e7e-68f7-4ba0-9728-826625811d81","added_by":"auto","created_at":"2022-12-16 16:03:34","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":36087,"visible":true,"origin":"","legend":"\u003cp\u003eFeature names and coefficients in models for differential diagnosisbetween benign and malignant lesions (a) and further identification csPCa in all lesions (b)\u003c/p\u003e","description":"","filename":"Fig.3ab.png","url":"https://assets-eu.researchsquare.com/files/rs-2324823/v1/ff26d4e54a115602ea9db99f.png"},{"id":30434902,"identity":"74eefcc7-0f24-484e-901b-43c214724f31","added_by":"auto","created_at":"2022-12-16 16:03:34","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":57703,"visible":true,"origin":"","legend":"\u003cp\u003eThe ROC curve of four model`s in predicting any tumor in PI-RADS 3 lesions. (a) training set, (b) internal test set, (c) external validation set\u003c/p\u003e","description":"","filename":"Fig.4abc.png","url":"https://assets-eu.researchsquare.com/files/rs-2324823/v1/eae3daa7ba05cbd3f9c6c3eb.png"},{"id":30435998,"identity":"768554fd-ba04-416e-b8b2-4fb810245d56","added_by":"auto","created_at":"2022-12-16 16:11:34","extension":"png","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":48212,"visible":true,"origin":"","legend":"\u003cp\u003eThe ROC curve of four models in predicting csPCa in PI-RADS 3 lesions. (a) training set, (b) internal test set, (c) external validation set\u003c/p\u003e","description":"","filename":"Fig.5abc.png","url":"https://assets-eu.researchsquare.com/files/rs-2324823/v1/ad1aa23b14afde5bacfa6080.png"},{"id":30435999,"identity":"90cf4431-48a5-4c7a-9539-ad8bd3a34483","added_by":"auto","created_at":"2022-12-16 16:11:34","extension":"png","order_by":6,"title":"Figure 6","display":"","copyAsset":false,"role":"figure","size":34884,"visible":true,"origin":"","legend":"\u003cp\u003eThe calibration plots of joint model in predicting all PCa (a, b) and csPCa (c, d) in PI-RADS 3 lesions. (a, c)internal test set, (b, d) external validation set\u003c/p\u003e","description":"","filename":"Fig.6abcd.png","url":"https://assets-eu.researchsquare.com/files/rs-2324823/v1/3553a5f33637cf27dd441bc9.png"},{"id":44724170,"identity":"e2e0ecc0-c216-4a6e-8155-ff325871020a","added_by":"auto","created_at":"2023-10-16 20:24:40","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":938248,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-2324823/v1/3f99bcc3-ae0d-4571-8da9-99e4e4b48611.pdf"}],"financialInterests":"No competing interests reported.","formattedTitle":"Machine Learning-Based Radiomics Model to Predict Benign and Malignant PI-RADS v2.1 Category 3 lesions : A Retrospective Multi-center Study","fulltext":[{"header":"Introduction","content":"\u003cp\u003eProstate cancer (PCa) is a global public health problem that threatens human health and life, which causes great harm to the male genitourinary system[\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e]. According to statistics from the American Cancer Research Association and the National Cancer Institute in 2019, PCa has become one of the most common malignant tumors in the world, accounting for the second most common malignancy in men[\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e]. MRI imaging is a common and highly effective exam for the diagnosis of prostate diseases. PI-RADS v2.1 published by American College of Radiology in 2019, represents a standardized method for assessing and reporting prostate MRI, which categorizes prostate lesions into different classes to reflect their relative likelihood of clinically significant prostate cancer (csPCa)[\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e]. PI-RADS 3 lesions included benign lesions and malignant lesions with different invasiveness and due to the absence of a clear tendency diagnosis for PI-RADS 3 lesions, there is a great variability in the practice patterns of different institutions (from conservative treatment, imaging follow-up to targeted biopsy), expense and potential clinical results[\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e]. Studies on evaluating the possibility of csPCa in targeted biopsies of PI-RADS 3 lesions have reported that cancer diagnosis rates range from 5\u0026ndash;30%, and most studies have suggested that the likelihood of eventual diagnosis of csPCa is relatively low[\u003cspan additionalcitationids=\"CR6\" citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e]. Therefore, accurately judging the benign and malignant lesions is helpful to reduce the pain caused by unnecessary biopsies.\u003c/p\u003e \u003cp\u003eImaging monitoring without intervention for PI-RADS category 3 lesions will undoubtedly reduce unnecessary biopsies. However, this method may lead to omission or delay in the diagnosis of csPCa lesions, resulting in irreversible consequences for patients. There is still controversy over whether to intervene in this category of lesions[\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e], and the small but not insignificant proportion of lesions that represent csPCa, it is critical that a more detailed classification of the category 3 lesions will benefit patients from biopsies and more aggressive treatment. Radiomics can convert images to higher-dimensional data, extract a large number of phenotypic features, and evaluate the biological behavior of tumor noninvasively through machine learning (ML) algorithms. It has been widely used in the diagnosis, invasiveness evaluation and clinical decision-making of PCa[\u003cspan additionalcitationids=\"CR10\" citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e]. The number of radiomics studies foucus on PI-RADS 3 lesions is limited only two single-center studies have previously assessed the role of radiomics characteristics to detect cancer in these \u0026ldquo;equivocal lesions\u0026rdquo;. However, there are doubts about the universality and wide applicability of radiomics models in the absence of multi-institution trials. Therefore, the purpose of this study was to construct a ML model combining T2WI, DWI and ADC radiomics features through a multi-center retrospective case-control study to validate its performance in differentiating PI-RADS 3 lesions from benign to malignant and in further risk stratification.\u003c/p\u003e"},{"header":"Materials And Methods","content":"\u003cp\u003e\u003cstrong\u003eStudy population and image acquisition\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThis retrospective multi-agency study was approved by the ethics review committee of each participating institution and exempted from the need for informed consent of the patient. Four medical centers have signed data sharing agreements for data exchange (2021; Approval No. 262). All prostate MRI images from January 2018 to December 2019 were exported from each participating unit\u0026apos;s PACS system. We summarized the data of each hospital, and there were a total of 2259 cases. 96 cases were excluded due to incomplete image sequence and lack of pathological data, then the remaining 2163 cases were divided into two parts and graded according to PI-RADS v2.1 criteria[3] by two radiologists with 3 years of experience in prostate MRI diagnosis, who were blind to pathological findings when reading. While interpreting the images, two radiologists recorded the location of each lesion using the anatomical fan map recommend by PI-RADS v2.1 to correspond to the lesion described by the pathological results. At an interval of two weeks after the first score, the procedure was repeated by two readers and reviewed by a senior radiologist proficient in MRI diagnosis of the urinary system. When there was any dispute over the interpretion, the three physicians discussed it until consensus was reached. Of the 2163 cases with final score results, 876 cases (40.5%) had PI-RADS scores of 1 and 2, 792 cases (36.6%) had PI-RADS scores of 4 and 5, and the remaining 495 cases (22.9%) were conferer with PI-RADS score 3. Then, all PI-RADS category 3 cases were selected for analysis, of which 32 were excluded based on the following criteria: (1) PI-RADS category 3 lesions coexisted with other types of lesions; (2) prior to MRI examination, they had received intervention such as biopsy, surgery or hormone therapy; (3) lack of any clinical characteristics of the patient or poor image quality. Finally, 463 eligible patients were recruited and MRI images of each patient showed only one lesion.\u003c/p\u003e\n\u003cp\u003eAll patients were divided into two groups according to the supplier of scanning equipment. The first group included institutions 1-3 with a total of 383 patients, which were examined with 3.0T superconducting MRI scanner (MAGNETOM Skyra, Germany) and equipped with 8-channel phased array body coils to collect signals. The second group consists of institution 4, with a total of 80 patients using a Dutch Philips Ingenia 3.0T MR scanner, the receiving coil was a 32-channel body phased array coil. The scanning sequences included T1WI, axial T2WI (no fat-saturated), sagittal T2WI, DWI (b=100, 800, 1500, 2000 s/mm\u003csup\u003e2\u003c/sup\u003e) and / or dynamic contrast enhanced T1WI. According to the DWI images with high b value, the ADC value was calculated by extended single exponential fitting model. The case datas of the first group were randomly divided into training set (n = 268) and internal test set (n = 115) according to the proportion of 7:3. The second\u003c/p\u003e\n\u003cp\u003egroup of cases was used as an external validation set (n = 80) to evaluate the extensibility of the model. During DCE scanning, 15 to 20 slices were scanned once, the scanning time resolution was 5.8 s, 64 phases were scanned, and the scanning time was 7min. After the end of the third dynamic scanning phase, contrast agent gadolinium meglumine pentanoate was injected intravenously at the injection rate of 3 ml/s and the dose of 0.1 mmol/kg. MRI scan parameters are described in Table 1. Fig. 1 provides a flowchart that includes patient selection and case assignment.\u003cstrong\u003e\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eTargeted biopsy and histopathology\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eMRI-TRUS fusion targeted sample was performed with Hitachi real-time ultrasonic multi-image fusion navigation system (RVS), and the machine model was HIVISIONNoblus/TopicPath. The suspicious lesions were sampled by MRI-TRUS fusion biopsy and systematic puncture under the guidance of TRUS within 4 weeks after the MRI examination. Before the fusion biopsy, the original data of prostate MRI in DICOM format were introduced into the main body of RVS ultrasound. MRI images were fused with TRUS images after general anesthesia, and anatomical markers such as urethra orifica, urethra, mullerian or ejaculatory duct cyst were matched with MRI sagittal images on the same section. T2WI, DWI, or DCE images with significant abnormal signals were selected to mark the target lesions in the cross-sectional MRI, while the same ultrasound sites were labeled (convex array scan), and then switched to sagittal images to further confirm the synchronization of MRI and ultrasound. After confirming favourable synchronization of MRI-TRUS images, the sagittal plane of prostate was taken by TRUS, and the target lesion marked with \u0026quot;+\u0026quot; was found. Under the guidance of puncture stent, the 18G disposable puncture gun was used to insert needle through perineum and the puncture gun was fired close to the target center. Then, the axial plane scan was converted to confirm that the needle track enters the target. 2-4 needles were punctured for each suspicious focus. After the targeted puncture, 12-needle puncture was conducted through perineum under the guidance of TRUS, and the pathological specimens were marked in detail according to each partition and fixed with 10% formaldehyde for pathological examination.\u003c/p\u003e\n\u003cp\u003eThe pathological results were evaluated by urological pathologist independently of the results of MRI, and the location and boundary of the lesions were recorded to ensure that they correspond to the suspicious lesions on MRI maps. The grade grouping and Gleason score of the lesions were determined according to the 2014 ISUP guidelines. csPCa was defined as ISUP Class 2 or higher (Gleason=3+4 or higher), and pathological results with GS=3+3 (ISUP class 1) were defined as clinically insignificant PCa (ciPCa)[12].\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eMRI Image preprocessing and focus segmentation\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eSubsequent evaluation and analysis were performed based on axial T2WI, DWI (b=2000 s/mm\u003csup\u003e2\u003c/sup\u003e) and ADC sequences in our study. The target images were spatially matched to ensure that DWI and ADC have the same resolution, field of view (FOV) and orientation compared with T2WI by using Elastix software package (v.4.10, http://elastix.isi.uu.nl/index.php)[13]. Before the focus segmentation, the images were standardized to improve the texture recognition rate. The \u0026mu;\u0026plusmn;3\u0026sigma; method was used to normalize the image, remove the gray signal more than 3 \u0026sigma;, and quantify the gray level with 64 levels. Finally, the voxel size of 1mm \u0026times; 1mm \u0026times; 1mm was used for equal-view resampling. The resampling operation was done by the \u0026quot;Radiomics\u0026quot; package of R software. Lesion segmentation was performed jointly by two radiologist involved in image evaluation using ITK-SNAP 3.8.0 software (http://www.itksnap.org/). The two handlers drew the region of interest (ROI) layer by layer on T2WI sequence to get the volume of interest (VOI) of the tumor, then copy it to DWI and ADC images to ensure the consistency of VOI sketches in different sequences. After preprocessing, visually verified was performed by a professor with experience in prostate MRI diagnosis (more than 10 years) to ensure that the location and extent of the lesions shown on MRI strictly matched the corresponding pathological description.\u003cstrong\u003e\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eFeature extraction and stability evaluation\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe open source radiomics software FeAture Explorer (FAE v0.4.0) developed by Song et al was used to extract features from the VOI of each image sequence[14]. According to the 8 texture analysis methods provided by the software, a total of 2347 image features were extracted from ROI files: (1) 46 first-order gray statistics; (2) 38 shape-based features; (3) 70 Gray Level Co-occurrence Matrices (GLCM); (4) 20 Gray Level Run Length Matrices (GLRLM); (5) 42 Gray Level Size Zone Matrices (GLSZM); (6) 36 Gray Level Dependence Matrices (GLDM); and (7) 17 Neighborhood Gray Tone Difference Matrices (NGTDM). (8) the original images were transformed by Wavelet Transform, and 2078 wavelet features are extracted in three spatial directions. The repeatability of intra- and inter-observer of lesion segmentation was based on the repeatability of feature extraction. 30 patients were randomly selected and the clinical data were blinded. The two doctors performed VOI segmentation and feature extraction again. The intra- and inter-observer repeatability of feature extraction was evaluated by intergroup correlation coefficient (ICC). If the intra-group and inter-group correlation coefficient is greater than 0.75, it is considered that the ROI drawing has acceptable stability.\u003cstrong\u003e\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eFeature selection and classifier modeling\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eIn this study, we focus on two results: (1) distinguish any cancer diagnosis from benign diagnosis, (2) and further predict csPCa occurrence in all cases. In order to solve the problem of sample imbalance in the characteristic matrix, this study uses the synthetic minority oversampling technique (SMOTE) to oversample the unbalanced positive and negative samples in the training set to balance the data set[15]. The number of image features was much larger than the number of samples, which may increase the risk of overfitting. This risk was reduced by feature selection to reduce the number of features. In present study, Z-Score Normalization was first used to normalize the feature matrix, each feature vector was subtracted from the mean value and divided by the standard deviation to eliminate the order of magnitude otherness between different features. The image features with variance of 0 were eliminated, and then the data dimension was reduced to remove the redundant features with average Spearman absolute correlation coefficient \u0026ge; 0.9. After eliminated redundant features, the analysis of variance (ANOVA) algorithm was used to sort the features, and only the top 20 features were retained. These features with increments from 1 to 20 were then input into the support vector machine (SVM) classifier. For different sequences, T2WI, DWI and ADC feature matrices were modeled respectively (called T2WI-model, DWI-model and ADC-model), and then the features of the three sequences were combined for modeling analysis (call integrated model). While established models to identify csPCa, the features of the first group were re-integrated, and the features of benign lesions and ciPCa were divided into the same group, and also divided into training set (n=268) and internal test set (n=115) according to the proportion of 7:3, and the generalization ability of the model was verified on the external valitation set. All the experiments above were run in FeAtureExplorer.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eStatistical analysis\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eDemographic datas were compared by chi-square test and independent t-test. According to whether it conformed to the normal distribution, the quantitative data were expressed as average (\u0026plusmn; standard deviation) or median (quartile range). Prediction models were inspected on the internal test and external validation sets. The receiver operating characteristic (ROC) curve was analyzed and the area under the ROC curve (AUC) was quantified to evaluate their performance in distinguishing cancer from benign lesions. Hosmer-lemeshow test was used to evaluate the degree of fitting between the predicted results of the integrated model and the histopathological results, and drawn the calibration diagram to visually display the results. In order to evaluate the generalization ability of the model, the non-inferior test was used to check whether the AUC of the external validation set is not lower than that of in the internal test set. R software (version 4.1.0, www. Rproject. org) was uesd for non-inferiority testing, the predefined acceptable threshold value was set to 0.1. Through the non-inferiority test of each model, the P-value was obtained, when P \u0026lt; 0.05, it indicates that the model has good versatility.\u003c/p\u003e"},{"header":"Results","content":"\u003cp\u003eClinical characteristics included only age and prostate specific antigen (PSA). The mean age of patients was 62.6\u0026plusmn;8.2 years, and PSA was 9.5\u0026plusmn;6.9 ng/mL. Of the 463 PI-RADS v2.1 category 3 lesions, 311 (67.2%) were benign and 152 (32.8%) were PCa lesions, of which 11.2%(52/463) were ciPCa (ISUP grade 1), 21.6% (100/463) were csPCa (47 ISUP level 2, 20 ISUP Level 3, 23 ISUP Level 4, 10 ISUP Level 5). There was no difference in the distribution of PCa and csPCa between different institutions (\u003cem\u003eP\u003c/em\u003e=0.502, 0.173). From the 463 PI-RADS category 3 lesion, there were 216 peripheral zone lesions (46.7%) with 79 PCa (48 csPCa and 31 ciPCa) and 247 transition zone lesions (53.3%) with 73 PCa (52 csPCa and 21ciPCa). The patient\u0026apos;s demographic and clinical datas were shown in Table 2.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eIn the intra- and inter-observer consistency test, the intra-observer ICCs range was 0.77 to 0.90, and the inter-observer ICCs range was 0.80 to 0.87, indicated that the repeatability of feature extraction was fine. Spearman correlation test results of the top 20 features screened by ANOVA were represented by feature heat map (Fig. 2). While constructed the integrated model, 6 and 5 features were screened to distinguish benign from malignant lesions and to further identify csPCa in cancer lesions. The name of the features and the corresponding coefficient are shown in fig 3.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eThe accuracy of T2WI-model in identifying csPCa of PI-RADS 3 lesions was 0.774 and 0.763 in internal test and external validation set, respectively, and the mean AUC value was 0.717 (internal test AUC = 0.738, external valitation AUC = 0.695, \u003cem\u003eP\u003c/em\u003e = 0.547). The accuracy of the model in distinguishing benign and malignant PI-RADS 3 lesions in internal test and external validation set was 0.644 and 0.650, respectively, with mean AUC of 0.624.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eThe accuracy of DWI-model in identifying csPCa of PI-RADS 3 lesions was 0.730 and 0.813 in internal test and external validation set, respectively, and the mean AUC value was 0.658 (internal test AUC = 0.635, external validation AUC = 0.681, \u003cem\u003eP\u003c/em\u003e = 0.547). The accuracy of the model in distinguishing benign and malignant PI-RADS 3 lesions in internal test and external validation set was 0.730 and 0.638, respectively, with mean AUC of 0.655.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eThe accuracy of ADC-model in identifying csPCa of PI-RADS 3 lesions was 0.739 and 0.775 in internal test and external validation set, respectively, and the mean AUC value was 0.746 (internal test AUC = 0.767, external validation AUC = 0.724, \u003cem\u003eP\u003c/em\u003e = 0.547). The accuracy of the model in distinguishing benign and malignant PI-RADS 3 lesions in internal test and external validation set was 0.565 and 0.613, respectively, with mean AUC value of 0.645.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eThe integrated model based on three single-sequence radiomics features, and its accuracy in identifying csPCa was 0.748 in internal test set and 0.863 in external validation set. The mean AUC value was 0.803 (internal test AUC = 0.804, external validation AUC = 0.801, \u003cem\u003eP\u003c/em\u003e = 0.019). The accuracy of the model in distinguishing benign and malignant PI-RADS 3 lesions in internal test and external validation set was 0.748 and 0.762, respectively, with mean AUC of 0.763. The results of Hosmer-Lemeshow test showed that the prediction results of the integrated model for all PCa and csPCa in the internal test and the external validation set had a high coincidence rate with the observed risks (P = 0.073 vs. 0.082 for PCa; \u003cem\u003eP\u003c/em\u003e = 0.224 vs. 0.647 for csPCa, respectively).\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eThe results of each model for distinguishing benign and malignant PI-RADS 3 diseases are shown in Table 3, and the corresponding ROC curves are shown in Fig. 4. The effectiveness of each model in identifying csPCa is compared in Table 4, and the corresponding ROC curves are shown in fig. 5. The pathological calibration scatter plots of the prediction results of the integrated model are shown in Fig. 6.\u003c/p\u003e"},{"header":"Discussion","content":"\u003cp\u003eThis study shows that ML models based on T2WI, DWI and ADC radiomics features can achieve upper-moderate accuracy when predicting any cancer and csPCa in PI-RADS v2.1 3 lesions, and the performance of integrated model is better than that of all single-sequence models, which indicates that only based on the simplex radiomics feature may be limited in distinguishing significant tumors from benign or inert lesions, and the combination of multiple features is well complementary. However, it is worth noting that the performance of all models in predicting csPCa is better than that of models in predicting all cancers. Therefore, our results also show that the heterogeneity of csPCa is more obvious than that of ciPCa, and it is easier to be recognized in ML progress.\u003c/p\u003e \u003cp\u003eSeveral additional indicators have been introduced to predict the need for biopsy in patients with PI-RADS 3, including lesion size, prostate volume, ADC, PSA and PSA density, but the published results do not fully prove the relationship between these indicators and the risk of csPCa appearance[\u003cspan additionalcitationids=\"CR17 CR18\" citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e]. For example, quantitative ADC values can help detect carcinoma while avoiding biopsies that are negative[\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e]. Zhang et al showed that age, PSA density, lesion zone and ADC value were predictors of csPCa and PCa[\u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e]. However, another study showed that the difference of median ADC values in PI-RADS 3 lesions was not statistically significant[\u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e22\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eIn recent years, radiomics studies have mainly focused on tumor detection, prediction of PI-RADS score and Gleason grade, evaluation of tumor extra-capsular extension and therapeutic response, which have shown similar performance as PI-RADS[\u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e23\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e24\u003c/span\u003e]. However, there are few studies use radiomics to assisted diagnose PI-RADS 3 lesions, and lack multi-center studies to validate the generalization ability of the model. Our results show that the single-sequence model is less efficient in both internal test and external validation set, with the lowest mean AUC for T2WI radiomics features, which is similar to the results of Lim et al[\u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e25\u003c/span\u003e]. They constructed a model based on XGBoost algorithm to predict any cancer or csPCa in PI-RADS 3 lesions, and AUC performed by T2WI features for all types of tumor was 0.608 and 0.547 for csPCa, lower than 0.642 and 0.684 of ADC features. Hectors et al. reconfirmed that model with T2WI radiomics features had a low ability to diagnose csPCa (AUC\u0026thinsp;=\u0026thinsp;0.76)[\u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e26\u003c/span\u003e]. However, ADC and DWI characteristics were not included as controls in their study. Our results are lower than those of Hou et al[\u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e27\u003c/span\u003e], who extracted features from T2WI, DWI and ADC images, constructed a one-step ML model and a regression analysis model integrated radiomics score, and improved the risk stratification method for identifying csPCa in PI-RADS 3 lesions with AUC reached 0.74 to 0.89.\u003c/p\u003e \u003cp\u003eThere are several design differences between this study and previous studies, which may explain the conflicts in results with Hou and Hectors. In contrast to these studies, our study used MRI datas from two vendors in four medical units. Different MRI scanners are equipped with different software and hardware, and these differences mean that scanners may not obtain images with the same intensity distribution. Although we performed resampling and registration operations before model construction, there is still the possibility of affecting the performance and generalization ability of the model. For this reason, an independent external validation set was set up to evaluate the model's performance, and this group of cases was provided by a different supplier from the testing set[\u003cspan additionalcitationids=\"CR29\" citationid=\"CR28\" class=\"CitationRef\"\u003e28\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR30\" class=\"CitationRef\"\u003e30\u003c/span\u003e]. The models constructed by Hou et al and Hectors et al were trained and tested only in their respective institutions, which limited extensibility. For example, quantitative values of DWI and ADC may be affected by variabilities between different scanners, imaging parameters, and patients, which caused the repeatability controversial. The lower accuracy of our study may be due to the fact that datas from multiple centers were integrated together and the number of PCa contributed by each participating unit was different, leading to differences in the distribution of cases. In order to ensure consistency between the combined data set and the distribution of cases in a single center, Lim et al. conducted a subgroup analysis of larger disease-causing institutions, but was unable to confirm this conjecture. Our study used non-inferiority test to evaluate the model's generalization ability, which was not available in other studies. Although we failed to prove that the AUC of all single-sequence radiomics featuers in the external validation set was not lower than that in the internal test set (\u003cem\u003eP\u003c/em\u003e\u0026thinsp;\u0026gt;\u0026thinsp;0.05). However, the diagnostic accuracy and sensitivity of the integrated model in external validation set are higher than that of the internal test set, and the AUC in external validation set was not inferior to the AUC in internal test set (\u003cem\u003eP\u003c/em\u003e\u0026thinsp;\u0026lt;\u0026thinsp;0.05), indicating that the integrated model has certain generalization ability in different date sets. In addition, Ji et al constructed a comprehensive model combinie age, PSA and radiomics features, suggested that combin clinical features can improve the generalization ability of radiomics model[\u003cspan citationid=\"CR31\" class=\"CitationRef\"\u003e31\u003c/span\u003e]. Different reference standards may also be one of the reasons for the different results. In Hou et al. 's study, a subset of included lesions lacked pathological diagnosis, and the clinical significance of tumor foci was only inferred based on follow-up imaging results and/or PSA changes after empirical treatment. This limits the reliability of the model's results for predicting a subset of clinically significant cancers, some of which were misclassified when they could have been monitored closely[\u003cspan citationid=\"CR32\" class=\"CitationRef\"\u003e32\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eFor the single-sequence model, the features based on DWI/ADC sequence are better than that based on T2WI in distinguishing between benign and malignant lesions. This is consistent with the research of Hou et al. In anther similar study, the most important feature for detecting tumor in PI-RADS 3 lesions was based on ADC maps[\u003cspan citationid=\"CR33\" class=\"CitationRef\"\u003e33\u003c/span\u003e]. The changes of diffusion of water molecules in tissues were monitored by DWI images, reflecting the changes of cell volume and number in epithelium, stroma and luminal space[\u003cspan citationid=\"CR34\" class=\"CitationRef\"\u003e34\u003c/span\u003e]. PCa is high cellular tissue, which restricts the diffusion to some extent due to the blocking of the random movement of water molecules in the tumor. The degree of diffusion limitation is positively correlated with the tumor grade, invasiveness and stage[\u003cspan citationid=\"CR35\" class=\"CitationRef\"\u003e35\u003c/span\u003e]. ML can quantify subtle changes in the diffusion motion of water molecules in the DWI/ADC diagram, which makes DWI perform better than other sequences to evaluate the PCa.\u003c/p\u003e \u003cp\u003eThere are several limitations in this study. First, this is a retrospective case-control study with a relatively small sample size, especially for a small number of csPCa with uneven distribution between groups, may be at risk of over-fitting when training models, which limits the evaluation of the accuracy in predicting malignant tumor; second, the sketch of ROI was done by radiologists by hand, which is time-consuming and affected by inter-observer variability. Automatic ROI segmentation algorithm can be introduced to improve the speed in the future. Third, the location of the lesion, such as peripheral and transitional zones, or poorly defined areas, was not taken into account. Due to the differences between peripheral zones and transition zones, modeling for each region may affect model performance. Finally, it was not discussed whether the clinical factors combined with radiomics features can provide additional diagnostic value for PI-RADS 3 lesions.\u003c/p\u003e"},{"header":"Conclusion","content":"\u003cp\u003eThe radiomics feature-based mechine learning model achieved an encouraging performance in differentiating PI-RADS 3 lesions from benign to malignant and distinguishing significant or indolent tumors, which has certain application value to assist clinical decision making, and provides a new direction for the management of patients with controversial MRI diagnosis and helps to reduce unnecessary biopsies while improving the detection rate of csPCa.\u003c/p\u003e"},{"header":"Abbreviations","content":"\u003cp\u003e\u003cstrong\u003ePCa\u003c/strong\u003e prostate cancer; \u003cstrong\u003ePI-RADS\u003c/strong\u003e prostate imaging report and data system; \u003cstrong\u003ecsPCa\u0026nbsp;\u003c/strong\u003eclinically significant prostate cancer; \u003cstrong\u003eML\u003c/strong\u003e machine learning; \u003cstrong\u003eROI\u003c/strong\u003e region of interest; \u003cstrong\u003eVOI\u0026nbsp;\u003c/strong\u003evolume of interest; \u003cstrong\u003eROC\u003c/strong\u003e receiver operating characteristic; \u003cstrong\u003eAUC\u003c/strong\u003e area under curve\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003e\u003cstrong\u003eAcknowledgements\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe authors thank all those who helped us during the writing of this research. We also thank the Department of Urology and Pathology of the hospitals for their valuable help and feedback.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAuthor contributions\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003ePengfei Jin: manuscript drafting and revision, study concept and design, collection, assembly, interpretation of the data, and figure drawing.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eLiqin Yang:manuscript drafting and figure drawing.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eJunkang Shen, Ao Shen and Ji Zhang: data collection and interpretation of the data.\u003c/p\u003e\n\u003cp\u003eJie Bao and Ximing Wang: Manuscript writing and final approval of the manuscript.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eAll authors read and approved the final manuscript.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eFundings\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThis study was supported by the Special Program for Diagnosis and Treatment Technology of Clinical Key Diseases in Suzhou (LCZX202001), Gusu health talent project of Suzhou (GSWS2020003), Suzhou Key Laboratory of health information technology (SZS201818).\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAvailability of data and materials\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe datasets generated during this study are available from the corresponding author upon reasonable request.\u003cstrong\u003e\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eEthical approval and Consent to Participate\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThis study was under ethics approval of the First Afliated Hospital of Soochow University \u0026nbsp;(Approval No. 262; 2021). All procedures performed in studies involving human participants were in accordance with the ethical standards of the institutional and/or national research committee and with the 1964 Helsinki declaration and its later amendments or comparable ethical standards.\u003c/p\u003e\n\u003cp\u003eInformed consent was obtained from all individual participants included in the study.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eConsent for publication\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eNot applicable.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eCompeting interests\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eAll authors report no conficts of interest.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\n \u003cli\u003eTeoh JYC, Hirai HW, Ho JMW, et al. Global incidence of prostate cancer in developing and developed countries with changing age structures. PLoS One. 2019 Oct 24;14(10):e0221775. doi: 10.1371/journal.pone.0221775.\u003c/li\u003e\n \u003cli\u003eCulp MB, Soerjomataram I, Efstathiou JA. Recent Global Patterns in Prostate Cancer Incidence and Mortality Rates. Eur Urol. 2020 Jan;77(1):38-52. doi: 10.1016/j.eururo.2019.08.005.\u0026nbsp;\u003c/li\u003e\n \u003cli\u003eTurkbey B, Rosenkrantz AB, Haider MA, et al.Prostate Imaging Reporting and Data System Version 2.1: 2019 Update of Prostate Imaging Reporting and Data System Version 2. Eur Urol,2019,76(3):340-351. doi: 10.1016/j.eururo.2019.02.033.\u003c/li\u003e\n \u003cli\u003eMaggi M, Panebianco V, Mosca A, et al. Prostate Imaging Reporting and Data System 3 Category Cases at Multiparametric Magnetic Resonance for Prostate Cancer: A Systematic Review and Meta-analysis. Eur Urol Focus. 2020 May 15;6(3):463-478. doi: 10.1016/j.euf.2019.06.014.\u003c/li\u003e\n \u003cli\u003eLiddell H, Jyoti R, Haxhimolla HZ. mp-MRI Prostate Characterised PIRADS 3 Lesions are Associated with a Low Risk of Clinically Significant Prostate Cancer - A Retrospective Review of 92 Biopsied PIRADS 3 Lesions. Curr Urol. 2015 Jul;8(2):96-100. doi: 10.1159/000365697.\u003c/li\u003e\n \u003cli\u003eSchoots IG. MRI in early prostate cancer detection: how to manage indeterminate or equivocal PI-RADS 3 lesions? Transl Androl Urol. 2018 Feb;7(1):70-82. doi: 10.21037/tau.2017.12.31.\u003c/li\u003e\n \u003cli\u003eOtti VC, Miller C, Powell RJ, et al. The diagnostic accuracy of multiparametric magnetic resonance imaging before biopsy in the detection of prostate cancer. BJU Int. 2019 Jan;123(1):82-90. doi: 10.1111/bju.14420.\u003c/li\u003e\n \u003cli\u003eLim CS, Abreu-Gomez J, Leblond MA, et al. When to biopsy Prostate Imaging and Data Reporting System version 2 (PI-RADSv2) assessment category 3 lesions? Use of clinical and imaging variables to predict cancer diagnosis at targeted biopsy. Can Urol Assoc J. 2021 Apr;15(4):115-121. doi: 10.5489/cuaj.6781.\u003c/li\u003e\n \u003cli\u003eKhalvati F, Zhang J, Chung AG, et al. MPCaD: a multi-scale radiomics-driven framework for automated prostate cancer localization and detection. BMC Med Imaging. 2018,18(1):16. doi: 10.1186/s12880-018-0258-4.\u003c/li\u003e\n \u003cli\u003eWang J, Wu CJ, Bao ML, et al. Machine learning-based analysis of MR radiomics can help to improve the diagnostic performance of PI-RADS v2 in clinically relevant prostate cancer. Eur Radiol. 2017 Oct;27(10):4082-4090. doi: 10.1007/s00330-017-4800-5.\u003c/li\u003e\n \u003cli\u003eMin X, Li M, Dong D, et al. Multi-parametric MRI-based radiomics signature for discriminating between clinically significant and insignificant prostate cancer: Cross-validation of a machine learning method. Eur J Radiol. 2019 Jun;115:16-21. doi: 10.1016/j.ejrad.2019.03.010.\u003c/li\u003e\n \u003cli\u003eEpstein JI, Egevad L, Amin MB, et al. The 2014 International Society of Urological Pathology (ISUP) Consensus Conference on Gleason Grading of Prostatic Carcinoma: Definition of Grading Patterns and Proposal for a New Grading System. Am J Surg Pathol. 2016 Feb;40(2):244-52. doi: 10.1097/PAS.0000000000000530.\u003c/li\u003e\n \u003cli\u003eKlein S, Staring M, Murphy K, et al. elastix: a toolbox for intensity-based medical image registration. IEEE Trans Med Imaging. 2010 Jan;29(1):196-205. doi: 10.1109/TMI.2009.2035616.\u003c/li\u003e\n \u003cli\u003eSong Y, Zhang J, Zhang YD, et al. FeAture Explorer (FAE): A tool for developing and comparing radiomics models. PLoS One. 2020 Aug 17;15(8):e0237587. doi: 10.1371/journal.pone.0237587.\u003c/li\u003e\n \u003cli\u003eSeo JH, Kim YH. Machine-Learning Approach to Optimize SMOTE Ratio in Class Imbalance Dataset for Intrusion Detection. Comput Intell Neurosci. 2018 Nov 1;2018:9704672. doi: 10.1155/2018/9704672.\u003c/li\u003e\n \u003cli\u003eFelker ER, Raman SS, Margolis DJ, et al. Risk Stratification Among Men With Prostate Imaging Reporting and Data System version 2 Category 3 Transition Zone Lesions: Is Biopsy Always Necessary? AJR Am J Roentgenol. 2017 Dec;209(6):1272-1277. doi: 10.2214/AJR.17.18008.\u003c/li\u003e\n \u003cli\u003eWashino S, Okochi T, Saito K, et al. Combination of prostate imaging reporting and data system (PI-RADS) score and prostate-specific antigen (PSA) density predicts biopsy outcome in prostate biopsy na\u0026iuml;ve patients. BJU Int. 2017 Feb;119(2):225-233. doi: 10.1111/bju.13465.\u003c/li\u003e\n \u003cli\u003eHermie I, Van Besien J, De Visschere P, et al. Which clinical and radiological characteristics can predict clinically significant prostate cancer in PI-RADS 3 lesions? A retrospective study in a high-volume academic center. Eur J Radiol. 2019 May;114:92-98. doi: 10.1016/j.ejrad.2019.02.031.\u003c/li\u003e\n \u003cli\u003eYang S, Zhao W, Tan S, et al. Combining clinical and MRI data to manage PI-RADS 3 lesions and reduce excessive biopsy. Transl Androl Urol. 2020 Jun;9(3):1252-1261. doi: 10.21037/tau-19-755.\u003c/li\u003e\n \u003cli\u003eGirometti R, Giannarini G, Panebianco V, et al. Comparison of different thresholds of PSA density for risk stratification of PI-RADSv2.1 categories on prostate MRI. Br J Radiol. 2021 Nov 11:20210886. doi: 10.1259/bjr.20210886.\u003c/li\u003e\n \u003cli\u003eZhang Y, Zeng N, Zhang FB, et al. Performing Precise Biopsy in Naive Patients With Equivocal PI-RADS, Version 2, Score 3, Lesions: An MRI-based Nomogram to Avoid Unnecessary Surgical Intervention. Clin Genitourin Cancer. 2020 Oct;18(5):367-377. doi: 10.1016/j.clgc.2019.11.011.\u003c/li\u003e\n \u003cli\u003eHermie I, Van Besien J, De Visschere P, et al. Which clinical and radiological characteristics can predict clinically significant prostate cancer in PI-RADS 3 lesions? A retrospective study in a high-volume academic center. Eur J Radiol. 2019 May;114:92-98. doi: 10.1016/j.ejrad.2019.02.031.\u003c/li\u003e\n \u003cli\u003eChaddad A, Niazi T, Probst S, et al. Predicting Gleason Score of Prostate Cancer Patients Using Radiomic Analysis. Front Oncol. 2018 Dec 18;8:630. doi: 10.3389/fonc.2018.00630.\u003c/li\u003e\n \u003cli\u003eGong L, Xu M, Fang M, Zou J, et al. Noninvasive Prediction of High-Grade Prostate Cancer via Biparametric MRI Radiomics. J Magn Reson Imaging. 2020 Oct;52(4):1102-1109. doi: 10.1002/jmri.27132.\u003c/li\u003e\n \u003cli\u003eLim CS, Abreu-Gomez J, Thornhill R, et al. Utility of machine learning of apparent diffusion coefficient (ADC) and T2-weighted (T2W) radiomic features in PI-RADS version 2.1 category 3 lesions to predict prostate cancer diagnosis. Abdom Radiol (NY). 2021 Dec;46(12):5647-5658. doi: 10.1007/s00261-021-03235-0.\u003c/li\u003e\n \u003cli\u003eHectors SJ, Chen C, Chen J, et al. Magnetic Resonance Imaging Radiomics-Based Machine Learning Prediction of Clinically Significant Prostate Cancer in Equivocal PI-RADS 3 Lesions. J Magn Reson Imaging. 2021 Nov;54(5):1466-1473. doi: 10.1002/jmri.27692.\u003c/li\u003e\n \u003cli\u003eHou Y, Bao ML, Wu CJ, et al. A radiomics machine learning-based redefining score robustly identifies clinically significant prostate cancer in equivocal PI-RADS score 3 lesions. Abdom Radiol (NY). 2020 Dec;45(12):4223-4234. doi: 10.1007/s00261-020-02678-1.\u003c/li\u003e\n \u003cli\u003eLitjens GJ, Hambrock T, Hulsbergen-van de Kaa C, et al. Interpatient variation in normal peripheral zone apparent diffusion coefficient: effect on the prediction of prostate cancer aggressiveness. Radiology. 2012 Oct;265(1):260-6. doi: 10.1148/radiol.12112374.\u003c/li\u003e\n \u003cli\u003eBarrett T, Lawrence EM, Priest AN, et al. Repeatability of diffusion-weighted MRI of the prostate using whole lesion ADC values, skew and histogram analysis. Eur J Radiol. 2019 Jan;110:22-29. doi: 10.1016/j.ejrad.2018.11.014.\u003c/li\u003e\n \u003cli\u003eSchmeel FC. Variability in quantitative diffusion-weighted MR imaging (DWI) across different scanners and imaging sites: is there a potential consensus that can help reducing the limits of expected bias? Eur Radiol. 2019 May;29(5):2243-2245. doi: 10.1007/s00330-018-5866-4.\u003c/li\u003e\n \u003cli\u003eJi X, Zhang J, Shi W, et al. Bi-parametric magnetic resonance imaging based radiomics for the identification of benign and malignant prostate lesions: cross-vendor validation. Phys Eng Sci Med. 2021 Sep;44(3):745-754. doi: 10.1007/s13246-021-01022-1.\u003c/li\u003e\n \u003cli\u003eBalakrishnan AS, Cowan JE, et al. Evaluating the Safety of Active Surveillance: Outcomes of Deferred Radical Prostatectomy after an Initial Period of Surveillance. J Urol. 2019 Sep;202(3):506-510. doi: 10.1097/JU.0000000000000247.\u003c/li\u003e\n \u003cli\u003eZhang KS, Schelb P, Kohl S, et al. Improvement of PI-RADS-dependent prostate cancer classification by quantitative image assessment using radiomics or mean ADC. Magn Reson Imaging. 2021 Oct;82:9-17. doi: 10.1016/j.mri.2021.06.013.\u003c/li\u003e\n \u003cli\u003eSurov A, Meyer HJ, Wienke A. Correlations between Apparent Diffusion Coefficient and Gleason Score in Prostate Cancer: A Systematic Review. Eur Urol Oncol. 2020 Aug;3(4):489-497. doi: 10.1016/j.euo.2018.12.006.\u003c/li\u003e\n \u003cli\u003eBrancato V, Aiello M, Basso L, et al. Evaluation of a multiparametric MRI radiomic-based approach for stratification of equivocal PI-RADS 3 and upgraded PI-RADS 4 prostatic lesions. Sci Rep. 2021 Jan 12;11(1):643. doi: 10.1038/s41598-020-80749-5.\u003c/li\u003e\n\u003c/ol\u003e"},{"header":"Tables","content":"\u003cp\u003e\u003cstrong\u003eTable 1 MRI protocols for both vender\u003c/strong\u003e\u003c/p\u003e\n\u003ctable align=\"\" border=\"0\" cellpadding=\"0\" cellspacing=\"0\" width=\"750\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd rowspan=\"2\" valign=\"top\" width=\"17.733333333333334%\"\u003e\n \u003cp\u003eMRI vendor Sequence\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"39.06666666666667%\"\u003e\n \u003cp\u003eSiemens Skyra 3.0T MR scanner (Germany)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"43.2%\"\u003e\n \u003cp\u003ePhilips Ingenia 3.0 T MR scanner (Netherlands)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" width=\"47.487844408427875%\"\u003e\n \u003cp\u003eT1WI \u0026nbsp; \u0026nbsp;AxialT2WI \u0026nbsp;SagittlT2WI \u0026nbsp; \u0026nbsp; DWI\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"52.512155591572125%\"\u003e\n \u003cp\u003eT1WI \u0026nbsp; AxialT2WI \u0026nbsp; \u0026nbsp; SagittlT2WI \u0026nbsp;DWI\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" width=\"17.733333333333334%\"\u003e\n \u003cp\u003eTR(ms)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"39.06666666666667%\"\u003e\n \u003cp\u003e680.0 \u0026nbsp; \u0026nbsp; 6980.0 \u0026nbsp; \u0026nbsp; \u0026nbsp; 3900.0 \u0026nbsp; 5000.0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"43.2%\"\u003e\n \u003cp\u003e556.0 \u0026nbsp; \u0026nbsp;3000.0 \u0026nbsp; \u0026nbsp; \u0026nbsp; 4978.0 \u0026nbsp; 6000.0\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" width=\"17.733333333333334%\"\u003e\n \u003cp\u003eTE(ms)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"39.06666666666667%\"\u003e\n \u003cp\u003e13.00 \u0026nbsp; \u0026nbsp; 104.00 \u0026nbsp; \u0026nbsp; \u0026nbsp; 89.00 \u0026nbsp; \u0026nbsp;72.00\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"43.2%\"\u003e\n \u003cp\u003e8.00 \u0026nbsp; \u0026nbsp; 100.00 \u0026nbsp; \u0026nbsp; \u0026nbsp; 100.00 \u0026nbsp; \u0026nbsp;77.00\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" width=\"17.733333333333334%\"\u003e\n \u003cp\u003eSlice thickness(mm)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"39.06666666666667%\"\u003e\n \u003cp\u003e5.0 \u0026nbsp; \u0026nbsp; \u0026nbsp; 3.0 \u0026nbsp; \u0026nbsp; \u0026nbsp; 3.0 \u0026nbsp; \u0026nbsp; \u0026nbsp;3.0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"43.2%\"\u003e\n \u003cp\u003e5.0 \u0026nbsp; \u0026nbsp; \u0026nbsp; 3.0 \u0026nbsp; \u0026nbsp; \u0026nbsp; 1.5 \u0026nbsp; \u0026nbsp; \u0026nbsp;3.0\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" width=\"17.733333333333334%\"\u003e\n \u003cp\u003eSlice gap(mm)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"39.06666666666667%\"\u003e\n \u003cp\u003e0.50 \u0026nbsp; \u0026nbsp; \u0026nbsp;0.00 \u0026nbsp; \u0026nbsp; \u0026nbsp;0.45 \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp;0.00\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"43.2%\"\u003e\n \u003cp\u003e0.00 \u0026nbsp; \u0026nbsp; \u0026nbsp;0.00 \u0026nbsp; \u0026nbsp; \u0026nbsp;0.15 \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp;0.00\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" width=\"17.733333333333334%\"\u003e\n \u003cp\u003eMatrix\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"39.06666666666667%\"\u003e\n \u003cp\u003e384\u0026times;384 \u0026nbsp;384\u0026times;384 \u0026nbsp; 384\u0026times;384 \u0026nbsp; \u0026nbsp; 130\u0026times;130\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"43.2%\"\u003e\n \u003cp\u003e276\u0026times;406 \u0026nbsp;240\u0026times;161 \u0026nbsp; \u0026nbsp; \u0026nbsp;276\u0026times;238 \u0026nbsp;124\u0026times;121\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" width=\"17.733333333333334%\"\u003e\n \u003cp\u003eFOV(mm\u0026times;mm)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"39.06666666666667%\"\u003e\n \u003cp\u003e380\u0026times;380 \u0026nbsp;200\u0026times;200 \u0026nbsp; 200\u0026times;200 \u0026nbsp; \u0026nbsp; 288\u0026times;288\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"43.2%\"\u003e\n \u003cp\u003e249\u0026times;415 \u0026nbsp;220\u0026times;220 \u0026nbsp; \u0026nbsp; \u0026nbsp;240\u0026times;180 \u0026nbsp;220\u0026times;220\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" width=\"17.733333333333334%\"\u003e\n \u003cp\u003eNSA\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"39.06666666666667%\"\u003e\n \u003cp\u003e\u0026nbsp; \u0026nbsp;1 \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp;2 \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; 3 \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp;2\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"43.2%\"\u003e\n \u003cp\u003e\u0026nbsp; \u0026nbsp; \u0026nbsp; 1 \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; 3 \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp;2 \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp;2\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n\u003cp\u003eTR repetition time; TE echo time; NSA number of signal averaged; T1WI T1 weighted imaging; T2WI T2 weighted imaging; DWI Diffusion Weighted Imaging\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eTable 2a Patient profiles of subgroup for distinguish any cancer diagnosis from benign lesions\u003c/strong\u003e\u003c/p\u003e\n\u003ctable align=\"\" border=\"0\" cellpadding=\"0\" cellspacing=\"0\" width=\"617\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" width=\"25.60777957860616%\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"16.207455429497568%\"\u003e\n \u003cp\u003eTraining set\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"16.045380875202593%\"\u003e\n \u003cp\u003eInternal test set\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"9.724473257698541%\"\u003e\n \u003cp\u003eP-value\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"22.042139384116695%\"\u003e\n \u003cp\u003eExternal validation set\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"10.372771474878444%\"\u003e\n \u003cp\u003eP-value*\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" width=\"25.60777957860616%\"\u003e\n \u003cp\u003eAges\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"16.207455429497568%\"\u003e\n \u003cp\u003e66.5\u0026plusmn;10.4\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"16.045380875202593%\"\u003e\n \u003cp\u003e69.8\u0026plusmn;7.6\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"9.724473257698541%\"\u003e\n \u003cp\u003e0.108\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"22.042139384116695%\"\u003e\n \u003cp\u003e70.2\u0026plusmn;12.4\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"10.372771474878444%\"\u003e\n \u003cp\u003e0.257\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" width=\"25.60777957860616%\"\u003e\n \u003cp\u003ePSA\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"16.207455429497568%\"\u003e\n \u003cp\u003e13.3\u0026plusmn;11.8\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"16.045380875202593%\"\u003e\n \u003cp\u003e14.7\u0026plusmn;9.3\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"9.724473257698541%\"\u003e\n \u003cp\u003e0.429\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"22.042139384116695%\"\u003e\n \u003cp\u003e11.2\u0026plusmn;7.3\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"10.372771474878444%\"\u003e\n \u003cp\u003e0.315\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" width=\"25.60777957860616%\"\u003e\n \u003cp\u003eLesion type\u003c/p\u003e\n \u003cp\u003eBenign\u003c/p\u003e\n \u003cp\u003eMalignant (ciPCa+csPCa)\u003c/p\u003e\n \u003cp\u003eTotal\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"16.207455429497568%\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003cp\u003e180\u003c/p\u003e\n \u003cp\u003e88 (30+58)\u003c/p\u003e\n \u003cp\u003e268\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"16.045380875202593%\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003cp\u003e77\u003c/p\u003e\n \u003cp\u003e38 (13+25)\u003c/p\u003e\n \u003cp\u003e115\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"9.724473257698541%\"\u003e\n \u003cp\u003e-\u003c/p\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"22.042139384116695%\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003cp\u003e54\u003c/p\u003e\n \u003cp\u003e26 (9+17)\u003c/p\u003e\n \u003cp\u003e80\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"10.372771474878444%\"\u003e\n \u003cp\u003e-\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n\u003cp\u003ePSA Prostate Specific Antigen;\u0026nbsp;csPCa clinically significant prostate cancer;\u0026nbsp;ciPCa non-clinically significant prostate cancer\u003c/p\u003e\n\u003cp\u003e*The P values are derived from the comparison between training set, internal test set and external validation set \u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eTable 2b Patient profiles of each subgroup for predicting clinically significant prostate cancer\u003c/strong\u003e\u003c/p\u003e\n\u003ctable align=\"\" border=\"0\" cellpadding=\"0\" cellspacing=\"0\" width=\"620\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" width=\"26.33279483037157%\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"15.347334410339258%\"\u003e\n \u003cp\u003eTraining set\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"15.993537964458804%\"\u003e\n \u003cp\u003eInternal test set\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"9.693053311793214%\"\u003e\n \u003cp\u003eP-value\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"21.970920840064622%\"\u003e\n \u003cp\u003eExternal validation set\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"10.662358642972537%\"\u003e\n \u003cp\u003eP-value*\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" width=\"26.33279483037157%\"\u003e\n \u003cp\u003eAges\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"15.347334410339258%\"\u003e\n \u003cp\u003e68.1\u0026plusmn;13.5\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"15.993537964458804%\"\u003e\n \u003cp\u003e66.7\u0026plusmn;5.9\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"9.693053311793214%\"\u003e\n \u003cp\u003e0.612\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"21.970920840064622%\"\u003e\n \u003cp\u003e69.3\u0026plusmn;4.4\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"10.662358642972537%\"\u003e\n \u003cp\u003e0.316\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" width=\"26.33279483037157%\"\u003e\n \u003cp\u003ePSA\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"15.347334410339258%\"\u003e\n \u003cp\u003e13.9\u0026plusmn;12.4\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"15.993537964458804%\"\u003e\n \u003cp\u003e12.6\u0026plusmn;10.4\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"9.693053311793214%\"\u003e\n \u003cp\u003e0.167\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"21.970920840064622%\"\u003e\n \u003cp\u003e13.4\u0026plusmn;5.1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"10.662358642972537%\"\u003e\n \u003cp\u003e0.121\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" width=\"26.33279483037157%\"\u003e\n \u003cp\u003eLesion type\u003c/p\u003e\n \u003cp\u003enon-csPCa (Benign+ciPCa)\u003c/p\u003e\n \u003cp\u003ecsPCa\u003c/p\u003e\n \u003cp\u003eTotal\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"15.347334410339258%\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003cp\u003e210 (180+30)\u003c/p\u003e\n \u003cp\u003e58\u003c/p\u003e\n \u003cp\u003e268\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"15.993537964458804%\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003cp\u003e90 (77+13)\u003c/p\u003e\n \u003cp\u003e25\u003c/p\u003e\n \u003cp\u003e115\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"9.693053311793214%\"\u003e\n \u003cp\u003e-\u003c/p\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"21.970920840064622%\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003cp\u003e63 (54+9)\u003c/p\u003e\n \u003cp\u003e17\u003c/p\u003e\n \u003cp\u003e80\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"10.662358642972537%\"\u003e\n \u003cp\u003e-\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n\u003cp\u003ePSA Prostate Specific Antigen;\u0026nbsp;csPCa clinically significant prostate cancer;\u0026nbsp;ciPCa non-clinically significant prostate cancer\u003c/p\u003e\n\u003cp\u003e*The P values are derived from the comparison between training set, internal test set and external validation set\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eTable 3 The performance of each model for predicting any tumors in PI-RADS 3 lesions\u003c/strong\u003e\u003c/p\u003e\n\u003ctable align=\"\" border=\"0\" cellpadding=\"0\" cellspacing=\"0\" width=\"716\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd rowspan=\"2\" valign=\"top\" width=\"13.668061366806137%\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003cp\u003eModality\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"22.175732217573223%\"\u003e\n \u003cp\u003eTraining set\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"23.291492329149232%\"\u003e\n \u003cp\u003eInternal test set\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"22.87308228730823%\"\u003e\n \u003cp\u003eExternal validation set\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"10.739191073919107%\"\u003e\n \u003cp\u003eMean AUC*\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd rowspan=\"2\" valign=\"top\" width=\"7.252440725244073%\"\u003e\n \u003cp\u003eP value\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" width=\"28.04232804232804%\"\u003e\n \u003cp\u003eAUC \u0026nbsp;ACC \u0026nbsp;SEN \u0026nbsp; SPE\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"29.45326278659612%\"\u003e\n \u003cp\u003eAUC \u0026nbsp;ACC \u0026nbsp;SEN \u0026nbsp; SPE\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"28.924162257495592%\"\u003e\n \u003cp\u003eAUC \u0026nbsp;ACC \u0026nbsp;SEN \u0026nbsp; SPE\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"13.580246913580247%\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" width=\"13.668061366806137%\"\u003e\n \u003cp\u003eT2WI-model\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"22.175732217573223%\"\u003e\n \u003cp\u003e0.811 \u0026nbsp;0.784 \u0026nbsp;0.614 \u0026nbsp;0.867\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"23.291492329149232%\"\u003e\n \u003cp\u003e0.678 \u0026nbsp;0.644 \u0026nbsp;0.842 \u0026nbsp;0.545\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"22.87308228730823%\"\u003e\n \u003cp\u003e0.589 \u0026nbsp;0.650 \u0026nbsp;0.500 \u0026nbsp;0.722\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"10.739191073919107%\"\u003e\n \u003cp\u003e0.624\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"7.252440725244073%\"\u003e\n \u003cp\u003e0.547\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" width=\"13.668061366806137%\"\u003e\n \u003cp\u003eDWI-model \u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"22.175732217573223%\"\u003e\n \u003cp\u003e0.717 \u0026nbsp;0.735 \u0026nbsp;0.557 \u0026nbsp;0.822\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"23.291492329149232%\"\u003e\n \u003cp\u003e0.712 \u0026nbsp;0.730 \u0026nbsp;0.684 \u0026nbsp;0.753\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"22.87308228730823%\"\u003e\n \u003cp\u003e0.598 \u0026nbsp;0.638 \u0026nbsp;0.615 \u0026nbsp;0.648\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"10.739191073919107%\"\u003e\n \u003cp\u003e0.655\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"7.252440725244073%\"\u003e\n \u003cp\u003e0.437\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" width=\"13.668061366806137%\"\u003e\n \u003cp\u003eADC-model\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"22.175732217573223%\"\u003e\n \u003cp\u003e0.840 \u0026nbsp;0.780 \u0026nbsp;0.773 \u0026nbsp;0.783\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"23.291492329149232%\"\u003e\n \u003cp\u003e0.650 \u0026nbsp;0.565 \u0026nbsp;0.921 \u0026nbsp;0.390\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"22.87308228730823%\"\u003e\n \u003cp\u003e0.640 \u0026nbsp;0.613 \u0026nbsp;0.654 \u0026nbsp;0.593\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"10.739191073919107%\"\u003e\n \u003cp\u003e0.645\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"7.252440725244073%\"\u003e\n \u003cp\u003e0.848\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" width=\"13.668061366806137%\"\u003e\n \u003cp\u003eIntegrated-model\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"22.175732217573223%\"\u003e\n \u003cp\u003e0.855 \u0026nbsp;0.746 \u0026nbsp;0.921 \u0026nbsp;0.661\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"23.291492329149232%\"\u003e\n \u003cp\u003e0.801 \u0026nbsp;0.748 \u0026nbsp;0.763 \u0026nbsp;0.740\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"22.87308228730823%\"\u003e\n \u003cp\u003e0.754 \u0026nbsp;0.762 \u0026nbsp;0.846 \u0026nbsp;0.722\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"10.739191073919107%\"\u003e\n \u003cp\u003e0.763\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"7.252440725244073%\"\u003e\n \u003cp\u003e0.047\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n\u003cp\u003eT2WI T2 weighted imaging; DWI Diffusion Weighted Imaging; ADC Apparent Diffusion Coefficient; AUC\u0026nbsp;area under the receiver operating characteristic curve; ACC accuracy; SEN sensitivity; SPE specificity\u003c/p\u003e\n\u003cp\u003e*Mean AUC\u0026nbsp;= [AUC(Internal test set) + AUC(External validation set)]/2\u003c/p\u003e\n\u003cp\u003eThe P-values from the non-inferiority tests\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eTable 4 The performance of each model for predicting csPCa in all PI-RADS 3 lesions\u003c/strong\u003e\u003c/p\u003e\n\u003ctable align=\"\" border=\"0\" cellpadding=\"0\" cellspacing=\"0\" width=\"714\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd rowspan=\"2\" valign=\"top\" width=\"13.72549019607843%\"\u003e\n \u003cp\u003emodality\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"22.969187675070028%\"\u003e\n \u003cp\u003eTraining set\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"22.689075630252102%\"\u003e\n \u003cp\u003eInner test set\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"22.408963585434172%\"\u003e\n \u003cp\u003eExternal validation set\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"11.064425770308123%\"\u003e\n \u003cp\u003eMean AUC*\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd rowspan=\"2\" valign=\"top\" width=\"7.142857142857143%\"\u003e\n \u003cp\u003eP value\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" width=\"29.02654867256637%\"\u003e\n \u003cp\u003eAUC \u0026nbsp;ACC \u0026nbsp;SEN \u0026nbsp; SPE\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"28.672566371681416%\"\u003e\n \u003cp\u003eAUC \u0026nbsp;ACC \u0026nbsp;SEN \u0026nbsp; SPE\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"28.31858407079646%\"\u003e\n \u003cp\u003eAUC \u0026nbsp;ACC \u0026nbsp;SEN \u0026nbsp; SPE\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"13.982300884955752%\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" width=\"13.72549019607843%\"\u003e\n \u003cp\u003eT2WI-model\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"22.969187675070028%\"\u003e\n \u003cp\u003e0.740 \u0026nbsp;0.668 \u0026nbsp;0.793 \u0026nbsp;0.633\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"22.689075630252102%\"\u003e\n \u003cp\u003e0.738 \u0026nbsp;0.774 \u0026nbsp;0.680 \u0026nbsp;0.800\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"22.408963585434172%\"\u003e\n \u003cp\u003e0.695 \u0026nbsp;0.763 \u0026nbsp;0.708 \u0026nbsp;0.778\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"11.064425770308123%\"\u003e\n \u003cp\u003e0.717\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"7.142857142857143%\"\u003e\n \u003cp\u003e0.264\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" width=\"13.72549019607843%\"\u003e\n \u003cp\u003eDWI-model \u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"22.969187675070028%\"\u003e\n \u003cp\u003e0.798 \u0026nbsp;0.802 \u0026nbsp;0.690 \u0026nbsp;0.833\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"22.689075630252102%\"\u003e\n \u003cp\u003e0.635 \u0026nbsp;0.730 \u0026nbsp;0.440 \u0026nbsp;0.811\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"22.408963585434172%\"\u003e\n \u003cp\u003e0.681 \u0026nbsp;0.813 \u0026nbsp;0.471 \u0026nbsp;0.905\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"11.064425770308123%\"\u003e\n \u003cp\u003e0.658\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"7.142857142857143%\"\u003e\n \u003cp\u003e0.086\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" width=\"13.72549019607843%\"\u003e\n \u003cp\u003eADC-model\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"22.969187675070028%\"\u003e\n \u003cp\u003e0.805 \u0026nbsp;0.784 \u0026nbsp;0.655 \u0026nbsp;0.819\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"22.689075630252102%\"\u003e\n \u003cp\u003e0.767 \u0026nbsp;0.739 \u0026nbsp;0.760 \u0026nbsp;0.733\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"22.408963585434172%\"\u003e\n \u003cp\u003e0.724 \u0026nbsp;0.775 \u0026nbsp;0.588 \u0026nbsp;0.825\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"11.064425770308123%\"\u003e\n \u003cp\u003e0.746\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"7.142857142857143%\"\u003e\n \u003cp\u003e0.269\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" width=\"13.72549019607843%\"\u003e\n \u003cp\u003eIntegrated-model\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"22.969187675070028%\"\u003e\n \u003cp\u003e0.854 \u0026nbsp;0.828 \u0026nbsp;0.741 \u0026nbsp;0.852\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"22.689075630252102%\"\u003e\n \u003cp\u003e0.804 \u0026nbsp;0.748 \u0026nbsp;0.800 \u0026nbsp;0.733\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"22.408963585434172%\"\u003e\n \u003cp\u003e0.801 \u0026nbsp;0.863 \u0026nbsp;0.921 \u0026nbsp;0.647\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"11.064425770308123%\"\u003e\n \u003cp\u003e0.803\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"7.142857142857143%\"\u003e\n \u003cp\u003e0.019\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n\u003cp\u003ecsPCa\u0026nbsp;clinically significant prostate cancer; T2WI T1 weighted imaging; DWI Diffusion Weighted Imaging; ADC Apparent Diffusion Coefficient; AUC area under the receiver operating characteristic curve; ACC accuracy; SEN sensitivity; SPE specificity\u003c/p\u003e\n\u003cp\u003e*Mean AUC\u0026nbsp;= [AUC(Internal test set) + AUC(External validation set)]/2\u003c/p\u003e\n\u003cp\u003eThe P-values from the non-inferiority tests\u003c/p\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":false,"highlight":"","institution":"","isAcceptedByJournal":true,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"bmc-medical-imaging","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"bmim","sideBox":"Learn more about [BMC Medical Imaging](http://bmcmedimaging.biomedcentral.com/)","snPcode":"","submissionUrl":"https://www.editorialmanager.com/bmim/default.aspx","title":"BMC Medical Imaging","twitterHandle":"BMC_series","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"em","reportingPortfolio":"BMC Series","inReviewEnabled":true,"inReviewRevisionsEnabled":true},"keywords":"Radiomics, Clinically significant prostate cancer, PI-RADS 3, Machine learning","lastPublishedDoi":"10.21203/rs.3.rs-2324823/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-2324823/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003e\u003cstrong\u003ePurpose: \u003c/strong\u003eTo develop machine learning-based prediction models derive from different MRI sequences for distinction between benign and malignant PI-RADS 3 lesions before intervention, and to cross-institution validate the generalization ability of the models.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eMethods: \u003c/strong\u003eThe pre-biopsy MRI datas of 463 patients diagnosed as PI-RADS 3 lesions were collected from 4 medical institutions. 2347 radiomics features were extracted from the VOI of T2WI, DWI and ADC maps. The ANOVA feature ranking method and support vector machine (SVM) classifier were used to construct 3 single-sequence models and 1 integrated model combined with the features of three sequences. All the models were established in the training set and independently verified in the internal test and external validation set. The area under the receiver operating characteristic curve (AUC) was used to quantify the predictive performance of each model. Hosmer-lemeshow test was used to evaluate the degree of fitting between prediction probability and pathological results. Non-inferiority test was used to check generalization performance of the integrated model.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eResults: \u003c/strong\u003eT2WI-model with the mean AUC of 0.717 for predicting clinically significant prostate cancer (csPCa) (internal test AUC = 0.738 vs. external validation AUC = 0.695, \u003cem\u003eP \u003c/em\u003e= 0.264) and 0.624 for predicting all cancer (internal test AUC = 0.678 vs. external validation AUC = 0.589, \u003cem\u003eP\u003c/em\u003e = 0.547). DWI-model with the mean AUC of 0.658 for predicting csPCa (internal test AUC = 0.635 vs. external validation AUC = 0.681, \u003cem\u003eP \u003c/em\u003e= 0.086) and 0.655 for predicting all cancer (internal test AUC = 0.712 vs. external validation AUC = 0.598, \u003cem\u003eP\u003c/em\u003e = 0.437). ADC-model with the mean AUC of 0.746 for predicting csPCa (internal test AUC = 0.767 vs. external validation AUC = 0.724, \u003cem\u003eP\u003c/em\u003e = 0.269) and 0.645 for predicting all cancer (internal test AUC = 0.650 vs. external validation AUC = 0.640, \u003cem\u003eP\u003c/em\u003e = 0.848). Integrated model with the mean AUC of 0.803 for predicting csPCa (internal test AUC = 0.804 vs. external validation AUC = 0.801, \u003cem\u003eP\u003c/em\u003e = 0.019) and 0.763 for predicting all cancer (internal test AUC = 0.801 vs. external validation AUC = 0.754,\u003cem\u003e P\u003c/em\u003e = 0.047).\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eConclusions:\u003c/strong\u003e The radiomics model based on mechine learning has the potential to be a non-invasive tool to distinguish cancerous, noncancerous and csPCa in PI-RADS 3 lesions and the generalization ability between different date set.\u003c/p\u003e","manuscriptTitle":"Machine Learning-Based Radiomics Model to Predict Benign and Malignant PI-RADS v2.1 Category 3 lesions : A Retrospective Multi-center Study","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2022-12-16 16:03:29","doi":"10.21203/rs.3.rs-2324823/v1","editorialEvents":[{"type":"communityComments","content":0},{"type":"decision","content":"Major revision","date":"2023-01-12T09:06:13+00:00","index":"","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2022-12-22T10:27:47+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"d48fd4d3-4377-45b9-b208-0ccc0bf00cba","date":"2022-12-19T22:19:20+00:00","index":"hide","fulltext":""},{"type":"reviewersInvited","content":"","date":"2022-12-19T22:16:47+00:00","index":"","fulltext":""},{"type":"editorAssigned","content":"","date":"2022-12-15T06:31:24+00:00","index":"","fulltext":""},{"type":"editorInvited","content":"","date":"2022-12-14T06:40:09+00:00","index":"","fulltext":""},{"type":"checksComplete","content":"","date":"2022-12-14T06:37:32+00:00","index":"","fulltext":""},{"type":"submitted","content":"BMC Medical Imaging","date":"2022-11-29T10:35:48+00:00","index":"","fulltext":""}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"bmc-medical-imaging","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"bmim","sideBox":"Learn more about [BMC Medical Imaging](http://bmcmedimaging.biomedcentral.com/)","snPcode":"","submissionUrl":"https://www.editorialmanager.com/bmim/default.aspx","title":"BMC Medical Imaging","twitterHandle":"BMC_series","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"em","reportingPortfolio":"BMC Series","inReviewEnabled":true,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"653ea50e-9dac-4c04-94ec-ed2b3b72a499","owner":[],"postedDate":"December 16th, 2022","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"published-in-journal","subjectAreas":[],"tags":[],"updatedAt":"2023-10-16T20:23:41+00:00","versionOfRecord":{"articleIdentity":"rs-2324823","link":"https://doi.org/10.1186/s12880-023-01002-9","journal":{"identity":"bmc-medical-imaging","isVorOnly":false,"title":"BMC Medical Imaging"},"publishedOn":"2023-03-29 20:14:50","publishedOnDateReadable":"March 29th, 2023"},"versionCreatedAt":"2022-12-16 16:03:29","video":"","vorDoi":"10.1186/s12880-023-01002-9","vorDoiUrl":"https://doi.org/10.1186/s12880-023-01002-9","workflowStages":[]},"version":"v1","identity":"rs-2324823","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-2324823","identity":"rs-2324823","version":["v1"]},"buildId":"FbvkV6FR0MCFSLy54lSbu","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}