Machine learning prediction of pathological complete response to neoadjuvant chemotherapy with peritumoral breast tumor ultrasound radiomics: compare with intratumoral radiomics and clinicopathologic predictors

Machine learning prediction of pathological complete response to neoadjuvant chemotherapy with peritumoral breast tumor ultrasound radiomics: compare with intratumoral radiomics and clinicopathologic predictors | 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 Help Center Sign In Submit a Preprint Cite Share Download PDF Research Article Machine learning prediction of pathological complete response to neoadjuvant chemotherapy with peritumoral breast tumor ultrasound radiomics: compare with intratumoral radiomics and clinicopathologic predictors Jiejie Yao, Wei Zhou, Xiaohong Jia, Ying Zhu, Xiaosong Chen, Weiwei Zhan, and 1 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-4440501/v1 This work is licensed under a CC BY 4.0 License Status: Posted Version 1 posted You are reading this latest preprint version Abstract Purpose Noninvasive, accurate and novel approaches to predict patients who will achieve pathological complete response (pCR) after neoadjuvant chemotherapy (NAC) could assist precise treatment strategies. The aim of this study was to explore machine learning (ML)-based peritumoral ultrasound radiomics signature (PURS), compared with intratumoral radiomics (IURS) and clinicopathologic factors, for early prediction of pCR. Methods We analyzed 358 locally advanced breast cancer patients (250 in the training set and 108 in the test set), who accepted NAC and post NAC surgery at our institution. The PURS and IURS of baseline breast tumors were extracted by using 3D-slicer and PyRadiomics software. Five ML classifiers including linear discriminant analysis (LDA), support vector machine (SVM), random forest (RF), logistic regression (LR), and adaptive boosting (AdaBoost) were applied to construct radiomics models for the prediction of pCR. The performance of PURS, IURS models and clinicopathologic predictors were assessed with respect to sensitivity, specificity, accuracy and the areas under the curve (AUCs). Results For the PURS models, the RF classifier achieved better efficacy (AUC of 0.889) than LR (0.849), AdaBoost (0.823), SVM (0.746) and LDA (0.732) in the test set. For the IURS models, the RF classifier also obtained a maximum AUC of 0.931 than 0.920 (AdaBoost), 0.875 (LR), 0.825 (SVM), and 0.798 (LDA) in the test set. The RF-based PURS yielded higher predictive ability (AUC, 0.889; 95% CI: 0.814, 0.947) than clinicopathologic factors (AUC, 0.759; 95% CI: 0.657, 0.861; p < 0.05), but lower efficacy compared with IURS (AUC, 0.931; 95%CI: 0.865, 0.980; p < 0.05). Conclusion The peritumoral US radiomics, as a novel potential biomarker, may be a promising clinical approach to guide precise therapy decisions. Machine learning Peritumoral and intratumoral ultrasound radiomics Pathological complete response Neoadjuvant chemotherapy Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Figure 6 Introduction For patients with locally advanced breast cancer (LABC), neoadjuvant chemotherapy (NAC) has been a standard treatment strategy to downstage tumor, reduce metastasis, and improve the probability of breast-conserving surgery [ 1 ]. Ideally, it could infer a favorable disease-free and an improved overall survival when a pathological complete response (pCR) was achieved after NAC [ 2 ]. However, according to the previous reports, approximately 30% of the patients may not respond to NAC, or even experience disease progression due to the morphological heterogeneous of tumors, or the genetic mutation [ 2 , 3 ]. Genetic identification is expensive and not routinely performed in clinical practice. Histopathologic examination is notable for its invasive procedure, and inevitable complications such as pain, hematoma and infection. Also, for heterogeneous tumor with different responses to NAC, a small tissue provided by biopsy is apt to selection bias [ 2 – 4 ]. Noninvasive, accurate and novel approaches to predict patients who will achieve pCR, and those who will benefit less but suffer more toxicities from NAC are highly desired. Radiomics can extract a large number of high-dimensional data from traditional medical images, and has been applied to predict therapeutic response. However, most of previous radiomics studies based on MRI [ 5 , 6 ], or CT [ 7 ], or PET/CT images [ 8 ], only few studies utilized US radiomics for prediction [ 9 ]. Moreover, prior researchers mainly focused on parameters within intratumoral structure, while the considerations about peritumoral region, which has been described as a “reactive zone” surrounding the tumor were limited. Recently, researchers began to explore intratumoral and peritumoral radiomics to predict NAC effect based on MRI [ 10 ] or mammography [ 11 ]. As a convenient, low cost, and widely used diagnostic tool, US plays an important role in screening and evaluating tumor margins. However, to our knowledge, no prior study used peritumoral US radiomics signatures (PURS) for the prediction of pCR to NAC. Compared to previous radiomics studies, most of which used the least absolute shrinkage and selection operator (LASSO) regression to establish predictive nomogram [ 5 – 9 ], machine learning (ML) algorithms such as linear discriminant analysis (LDA), support vector machine (SVM), random forest (RF), logistic regression (LR) and adaptive boosting (Adaboost) can handle abundant quantitative radiomics features powerfully and effectively [ 12 ]. Nevertheless, the application of ML-based PURS and IURS to predict NAC effect has not been explored. A prior study of our team has constructed four ML classifiers-based US radiomics for preoperative prediction of axillary sentinel lymph node metastasis burden in early-stage BC patients [ 13 ]. Now, we were interested in whether PURS can be employed to predict pCR after NAC. We also wondered the predictive ability of various ML algorithms based radiomics models. Therefore, the purpose of this study was to assess the efficacy of five ML classifiers-based PURS, compared with IURS and clinicopathologic factors, for early prediction of pCR after NAC in LABC patients. Methods Study population This study was a retrospective analysis, and was conducted in accordance with the Declaration of Helsinki, approved by the Ethics Committee of our institution. The informed consent was waived due to its retrospective nature. Data sets were obtained from January 2018 to December 2022, we firstly enrolled 505 LABC patients who underwent pretreatment breast tumor biopsy, then accepted NAC and post NAC surgery at our institution. The inclusion criteria were as follows: (i) patients with biopsy-proven primary BC, and without distant metastasis; (ii) patients with high quality baseline breast tumor US images before biopsy; (iii) patients underwent a full course of NAC; (iv) patients accepted surgery after NAC, and pCR or non-pCR status was confirmed by surgical specimen histopathologic examinations. The exclusion criteria were as follows: (i) patients who did not complete NAC regimen; (ii) patients with bilateral BC or multiple tumors; (iii) patients with nonmass-like lesions; (iv) patients with no sufficient peritumoral tissue identified on US images. Finally, a total of 358 patients (all women, mean age, 48.1 years± 10.5; median age, 49.6 years; age range, 38–79 years) were included, and were divided into a training set (from January 2018 to December 2020, n = 250), and an independent test set (from January 2021 to December 2022, n = 108). Figure 1 shows the patient recruitment and study design. Data collection The clinical data such as patients’ age, clinical T and N stage, and NAC regimens were recorded and retrieved from the Shanghai Jiaotong University Breast Cancer Database (SJTU-BCDB). The pathological data included tumor histological type, tumor proliferation rate (Ki67 levels), estrogen receptor (ER), progesterone receptor (PR), and human epidermal growth factor receptor 2 (HER2) status. HER2 status was confirmed with fluorescence in situ hybridization. A cut-off value for Ki67 positive was established at 20% [ 14 ]. Tumors were classified into 3 subgroups based on the expression of ER, PR, and HER2 status: HR+/HER2- (HR+, HER2-); HER2+ (HER2+, ER + or ER-, PR + or PR-); and triple-negative (ER-, PR-, HER2-). HR + was defined as ER + and/or PR+. All the patients received six or eight cycles of NAC before breast surgery according to the National Comprehensive Cancer Network (NCCN) guideline [ 15 ]. The pCR is defined as no residual invasive (ductal carcinoma in situ could be present), and no axillary lymph node invasion in the final post NAC surgical specimen (ypT0/isN0) [ 16 ]. US images acquisition and segmentation Breast US examinations were performed 1 week before biopsy by Resona 7 (Mindray Medical International, Shenzhen, China) with a linear probe at 3–11 MHz, and Esaote MyLab 60 (Esaote, Genoa, Italy) with a linear probe at 4–13 MHz. Tumors were assessed according to the Breast Imaging Reporting and Data System (BI-RADS) [ 17 ]. The maximum size of the breast tumors measured by US were also recorded. The tumor regions of interest (ROIs) segmentation were performed via a free open-source software package (3D Slicer version 5.0.3) [ 18 ]. The intratumoral region was segmented by dilating the delineated tumor contour manually in the largest cross-sectional area. The peritumoral ROI was obtained with a 3mm-thick surrounding zone outside the intratumoral region automatically using “Hollow” and “Margin” segment editors. The ROIs of peritumoral and intratumoral area were extracted separately with PyRadiomics software [ 18 ]. Figure 2 (a-h) shows the examples of tumor US images and its corresponding ROIs. Intra- and interobserver reproducibility were assessed with intra- and interclass correlation coefficients (ICCs). Two experienced radiologists (author 1 and author 2, with over fifteen years of experience in breast US, and three years of experience in the software) who were blinded to the treatment outcomes, segmented the peritumoral and intratumoral regions of 60 randomly selected breast tumors, and extracted the radiomics signatures separately. One weeks later, author 1 repeated the same procedure and analyzed the remaining images. An ICC equal to or higher than 0.75 was considered as good intra- and interobserver agreement, and was included in the further feature selection process. Feature extraction, selection and ML-classifiers implementation For each breast tumor, a total of 851 radiomics features were extracted and categorized as follows: 14 shape features, 18 first-order features, 24 gray level co-occurrence matrix (GLCM), 14 gray level dependence matrix (GLDM), 16 gray level run length matrix (GLRLM), 16 gray level size zone matrix (GLSZM), 5 neighboring gray tone difference matrix (NGTDM), and 744 wavelet-related features (details are shown in Supplementary Appendix A and B). The final training sets comprised each of 212,750 PURS and IURS, the independent test sets contained each of 91,908 radiomics signatures. To select the most robust features, synthetic minority oversampling technique (SMOTE) was firstly used to remove the unbalance samples in the data set [ 19 ]. Then, Z-score and Mean normalization methods were applied to standardize the corresponding features. Principal component analysis (PCA) and pearson correlation coefficient (PCC) were employed to increase data interpretation and reduce feature dimension. After that, recursive features elimination (RFE) was used to detect the most relevant signatures [ 20 ]. Finally, the most significant selected features were input to LDA, SVM, RF, LR, and Adaboost classifiers with a 5-fold cross validation to construct PURS and IURS models for the prediction of pCR. Figure 3 shows the overview of the workflow. Statistical Analysis All numerical data were presented as mean ± standard deviation. Continuous and categorical variables were compared using the independent t test and the Chi-square test or Fisher’s exact test, respectively. All clinical and pathological factors that were shown to be potentially associated with pCR ( p < 0.05 in univariate analysis) were considered to construct a clinicopathologic model. The training data set was used to construct five ML classifiers-based PURS and IURS predictive models, the test data set was used for independent validation to evaluate the performance of the models. The predictive abilities were assessed with respect to sensitivity (SEN), specificity (SPE), accuracy (ACC), positive predictive value (PPV), negative predictive value (NPV), and the area under the receiver operating characteristic (ROC) curve (AUC). Comparisons between AUCs were made by using the DeLong test. All of the processes were implemented with FeAture Explorer Pro (FAEPro, V0.5.3) in Python (3.7.6) [ 21 ] ( https://github.com/salan668/FAE ) , SPSS software (version 23.0), and MedCalc software (version 22.013). A P value less than 0.05 was regarded as statistically significant difference. Results Clinicopathologic characteristics Among all the 358 patients, the rate of pCR was 27.1% (97/358). It was 27.2% (68/250) in the training set, and 26.9% (29/108) in the test set. The ER, PR, and HER2 status were found to be significantly associated with pCR in both sets ( p 0.05 for all). Regarding molecular subtypes, the rate of pCR was significantly higher in HER2 + patients [58.8% (40/68), 69.0% (20/29)], than triple-negative [25.0% (17/68), 17.2% (5/29)], and HR+/HER2- patients [16.2% (11/68), 13.8% (4/29)] in the training and test sets, respectively ( p < 0.001 for all) (Table 1 ). Table 1 Clinicopathologic data of patients in relation to pCR and non-pCR status in the training and test sets Training set (n = 250) No. (%) Test set (n = 108) No. (%) Characteristics (n = 68) pCR (n = 182) Non-pCR P value (n = 29) pCR (n = 79) Non-pCR P value Age, years 47.2 ± 10.6 50.2 ± 11.9 0.346 48.5 ± 12.3 51.1 ± 13.0 0.217 Tumor maximum size (mm) 37.4 ± 11.6 38.5 ± 13.4 0.659 38.2 ± 12.8 39.0 ± 14.3 0.588 Clinical T stage 0.334 0.512 I 5 (7.4) 7 (3.8) 2 (6.9) 3 (3.8) II 36 (52.9) 88 (48.4) 16 (55.2) 37 (46.8) III 27 (39.7) 87 (47.8) 11 (37.9) 39 (49.4) Clinical N stage 0.273 0.392 cN0 16 (23.5) 30 (16.5) 6 (20.7) 11 (13.9) cN+ 52 (76.5) 152 (83.5) 23 (79.3) 68 (86.1) Histologic type 0.497 0.685 Invasive ductal carcinoma 49 (72.1) 123 (67.6) 19 (65.5) 55 (69.6) Others 19 (27.9) 59 (32.4) 10 (34.5) 24 (30.4) ER status < 0.001* 0.002* Positive 26 (38.2) 141 (77.5) 10 (34.5) 56 (70.9) Negative 42 (61.8) 41 (22.5) 19 (65.5) 23 (29.1) PR status < 0.001* < 0.001* Positive 22 (32.4) 125 (68.7) 8 (27.6) 55 (69.6) Negative 46 (67.6) 57 (31.3) 21 (72.4) 24 (30.4) HER2 status < 0.001* 20% 55 (80.9) 139 (76.4) 25 (86.2) 62 (78.5) Molecular subtypes < 0.001* < 0.001* HR+/HER2- 11 (16.2) 90 (49.5) 4 (13.8) 45 (56.9) HER2+ 40 (58.8) 59 (32.4) 20 (69.0) 21 (26.6) Triple-negative 17 (25.0) 33 (18.1) 5 (17.2) 13 (16.5) Note.—There are 358 patients in the training and test sets. Mean data are ± standard deviation with the range. Data are the number of patients, with percentages in parentheses. ER = estrogen receptor, PR = progesterone receptor, HER2 = human epidermal growth factor receptor 2, HR = hormone receptor, pCR = pathological complete response; * p value < 0.05. The clinicopathologic model based on the factors associated with pCR in univariate analysis (i.e., ER, PR, and HER2 status, and molecular subtypes) yielded an AUC of 0.776 (95% CI: 0.717, 0.835), SEN of 78.6%, SPE of 82.1%, ACC of 76.9%, PPV of 68.3%, and NPV of 85.6% in the training set. In the test set, it obtained an AUC of 0.759 (95% CI: 0.657, 0.861), SEN of 74.6%, SPE of 79.3%, ACC of 74.8%, PPV of 66.5%, and NPV of 84.1%. The inter- and intraobserver reproducibility were substantial for the ROIs segmentation and radiomics features extraction, with ICCs > 0.75 for all, and were robust for the further analysis. Performance of ML-based PURS and IURS models For the PURS models, the RF classifier achieved the best predictive ability with an AUC of 0.882 (95% CI: 0.838, 0.923), than LR (AUC = 0.866 [95% CI: 0.799, 0.921]), Adaboost (AUC = 0.834 [95% CI: 0.760, 0.901]), SVM (AUC = 0.738 [95% CI: 0.664, 0.819]), and LDA (AUC = 0.687 [95% CI: 0.617, 0.761]) in the training set. In the test set, the AUCs were 0.889 (95% CI: 0.814, 0.947) for RF, 0.849 (95% CI: 0.734, 0.942) for LR, 0.823 (95% CI: 0.697, 0.933) for Adaboost, 0.746 (95% CI: 0.629, 0.848) for SVM, and 0.732 (95% CI: 0.625, 0.835) for LDA (Fig. 4 a-e). The RF-based PURS predictive model also yielded favorable SEN of 80.8%, SPE of 87.4%, ACC of 85.6%, PPV of 70.5%, and NPV of 92.4% in the training set, and SEN of 75.9%, SPE of 89.7%, ACC of 85.9%, PPV of 73.3%, and NPV of 90.9% in the test set, respectively (Table 2 ). Table 2 The predictive performance of ML-based PURS, IURS models in the training and test sets Training set SEN (%) SPE (%) ACC (%) PPV (%) NPV (%) AUC (95% CI) PURS models RF 80.8 87.4 85.6 70.5 92.4 0.882 (0.838–0.923) LR 77.9 93.4 89.2 81.5 91.9 0.866 (0.799–0.921) Adaboost 77.8 91.8 88.0 77.9 90.3 0.834 (0.760–0.901) SVM 80.4 66.7 72.5 49.9 93.8 0.738 (0.664–0.819) LDA 69.1 69.4 69.3 45.6 85.8 0.687 (0.617–0.761) IURS models RF 92.6 95.6 93.2 88.7 97.2 0.948 (0.913–0.978) Adaboost 89.7 77.6 81.8 63.8 95.3 0.910 (0.873–0.943) LR 73.5 81.4 79.3 59.6 89.3 0.852 (0.800-0.897) LDA 67.7 81.9 80.7 58.3 87.2 0.825 (0.757–0.874) SVM 72.3 74.5 74.3 59.2 85.6 0.817 (0.739–0.868) Test set SEN (%) SPE (%) ACC (%) PPV (%) NPV (%) AUC (95% CI) PURS models RF 75.9 89.7 85.9 73.3 90.9 0.889 (0.814–0.947) LR 79.3 93.5 87.7 82.1 92.4 0.849 (0.734–0.942) Adaboost 75.8 96.1 90.6 80.1 91.4 0.823 (0.697–0.933) SVM 68.9 85.9 81.3 64.5 88.1 0.746 (0.629–0.848) LDA 72.4 73.0 72.9 50.0 87.6 0.732 (0.625–0.835) IURS models RF 96.5 93.5 94.3 88.8 98.6 0.931 (0.865–0.980) Adaboost 89.6 79.5 84.4 88.4 95.3 0.920 (0.869–0.967) LR 79.3 84.6 83.2 65.7 91.6 0.875 (0.787–0.935) SVM 72.4 84.4 81.3 64.9 89.1 0.825 (0.734–0.904) LDA 75.0 83.3 72.6 63.7 88.2 0.798 (0.699–0.887) Note. — SEN = sensitivity, SPE = specificity, ACC = accuracy, PPV = positive predictive value, NPV = negative predictive value, AUC = area under the receiver operating curve, CI = confidence interval, PURS = peritumoral ultrasound radiomics signature, IURS = intratumoral ultrasound radiomics signature, RF = random forest, AdaBoost = adaptive boosting, LR = logistic regression, SVM = support vector machine, LDA = linear discriminant analysis. Ten, 7, 5, 3, and 2 optimal radiomics features were selected for the RF, LR, Adaboost, SVM, and LDA classifiers, respectively. The top three selected features in the RF classifier were wavelet-HHL_GLSZM_gray level non uniformity normalized (coefficient: 4.492), wavelet-LHH_GLRLM_run length non uniformity normalized (coefficient: 3.579), and wavelet-HLH_GLSZM_high gray level zone emphasis (coefficient: 1.700) (details are shown in Supplementary Appendix C-G). For the IURS models, the RF classifier obtained a maximum AUC of 0.948 (95% CI: 0.913, 0.978), compared with Adaboost (AUC = 0.910 [95% CI: 0.873, 0.943), LR (AUC = 0.852 [95% CI: 0.800, 0.897]), LDA (AUC = 0.825 [95% CI: 0.757, 0.874]), and SVM (AUC = 0.817 [95% CI: 0.739, 0.868]) in the training set. In the test set, the AUCs were 0.931(95% CI: 0.865, 0.980) for RF, 0.920 (95% CI: 0.869, 0.967) for Adaboost, 0.875 (95% CI: 0.787, 0.936) for LR, 0.825 (95% CI: 0.734, 0.904) for SVM, and 0.798 (95% CI: 0.699, 0.887) for LDA (Fig. 5 a-e). The RF-based IURS predictive model also achieved a satisfactory SEN of 92.6%, SPE of 95.6%, ACC of 93.2%, PPV of 88.7%, and NPV of 97.2% in the training set, and SEN of 96.5%, SPE of 93.5%, ACC of 94.3%, PPV of 88.8%, and NPV of 98.6% in the test set, respectively (Table 2 ). Eleven, 8, 6, 4 and 2 optimal radiomics features were selected for the RF, Adaboost, LR, SVM, and LDA classifiers, respectively. The top three selected features in the RF classifier were wavelet-LHH_GLDM_large dependence low gray level emphasis (coefficient: 6.192), wavelet-LLL_GLCM_correlation (coefficient: 4.935), and wavelet-HHL_GLRLM_short run emphasis (coefficient: 2.302) (details are shown in Supplementary Appendix H-L). DeLong test showed that both RF-based PURS and IURS models had higher ability than clinicopathologic model (0.882 vs. 0.776, Z = 3.017, p = 0.001, and 0.948 vs. 0.776, Z = 4.788, p < 0.001) in the training set, and (0.889 vs. 0.759, Z = 3.646, p < 0.001, and 0.931 vs. 0.759, Z = 4.059, p < 0.001) in the test set. Nevertheless, the RF-based PURS showed lower efficacy as compared with RF-based IURS (0.882 vs. 0.948, Z = 2.970, p = 0.003, and 0.889 vs. 0.931, Z = 2.247, p = 0.02) in both sets (Fig. 6 a and b). Discussion Reliable and noninvasive predictors of pCR may assist clinicians with precise NAC tactics for LABC patients. Previous researchers adopted breast cancer subtypes [ 22 ], US and mammographic images [ 23 ] to target NAC. More studies utilized MRI radiomics as a standardized image method to monitor treatment response [ 5 , 6 , 10 ]. A recent study reported that US radiomics compared favorably with MRI in the assessment of pCR [ 24 ]. However, it still remains unknown whether PURS may contribute to the response prediction. Also, the predictive value of various ML classifiers-based US radiomics has not been explored. The present study is the first attempt to assess the ability of various ML classifiers-based PURS, as compared with IURS and clinicopathologic factors to predict NAC effect. Our results showed that the RF classifier-based PURS exhibited a higher performance than clinicopathologic predictors, and a relatively low efficacy compare with IURS for the prediction of pCR ( p < 0.05 for both) . Peritumoral regions which contain a mixture of tumor cells and inflammatory elements have been reported to be associated with tumor aggressive and prognosis [ 25 ]. Braman et al. [ 10 ] applied intratumoral and peritumoral DCE-MRI radiomics for the pretreatment prediction of pCR to NAC. They defined 2.5- to 5mm radius surrounding the tumor as the peritumoral region, and yielded an AUC of 0.74 by using a combined set. Li et al. [ 26 ] reported that the ability to differentiate benign and malignant breast lesions within 3mm peripheral regions was similar to that of the entire internal regions on contrast-enhanced sonography (CEUS). Mao et al. [ 11 ] segmented five regions including intratumoral region, 5mm peritumoral region, 10mm peritumoral region, intratumoral + 5mm peritumoral regions, and intratumoral + 10mm peritumoral regions, and compared their performances on contrast-enhanced spectral mammography. Nevertheless, their study focused on the statistical difference between different areas of ROIs. Regions further away from the tumor boundary may involve more normal tissue, and embody less outcome-related information. In this study, we selected a 3mm-thick zone surrounding the tumor as the commonly used peritumoral region for PURS analysis, our results achieved a higher predictive efficacy than that of the combined models reported by Braman et al. [ 10 ]. However, as compared with IURS, the PURS models yielded a relatively lower predictive ability, which suggested that the immune response and the expansive lymphatic vessels with a relative preservation structure surrounding tumors may lead to resistance to treatment response. Radiomics feature havs been regard as a potential predictor for clinical outcomes. In the present study, the dominant features selected in both PURS and IURS models were wavelet-related features. After the wavelet transform, GLSZM and GLRLM were the most selected features in PURS models, GLCM and GLDM were the mainly features in IURS models. GLRLM contains the length of the grayscale value parade, provides information on the spatial distribution of consecutive pixels in one or more directions. GLSZM is based on a principle similar to GLRLM, and can be calculated for the distances of different pixels or regions in the neighborhood. These features can indicate the comprehensive changed information of images in adjacent regions [ 27 ]. GLCM is a symmetric matrix representing the joint probability distribution of pixel pairs. GLDM includes the emphasis on large and small dependencies representing heterogeneity and homogeneity. Previous studies revealed that these signatures reflected the texture roughness and unevenness of an image, and indicated the intrinsic tumor heterogeneity [ 28 ]. Wavelet transform provides more valuable information about the tumor micro-environment through the recollection of texture features, with higher details and complexities than the original images [ 27 , 28 ]. Wavelet transform features has been proved to improve the diagnostic ability of malignant, the classification of breast cancer, and the prediction of prognosis or metastatic behavior [ 29 , 30 ]. Our results showed that the wavelet-related features of breast tumors were reliable predictors associated with pCR after NAC, which was compatible with previous studies [ 28 , 29 ]. In addition to ML algorithms, we used five popular classifiers (i.e., LDA, SVM, RF, LR, and AdaBoost) to construct PURS and IURS predictive models. The RF classifier achieved the most robust performance as compared with other classifiers in both PURS and IURS models. Tahmassebi et al. [ 31 ] applied eight ML classifiers with multiparametric MRI to predict pCR and survival outcomes. Their result showed that the XGBoost classifier achieved the most high accuracy for predicting residual cancer burden and disease-specific survival, and LR yielded a high AUC of 0.83 for predicting recurrence-free survival. However, their studies limited with only 38 patients, the small sample may be insufficient for clinical application. Sun et al. [ 32 ] indicated that RF classifier obtained the highest ability in the detection of TP53 mutations in TN and luminal type BCs. Masetic et al. [ 33 ] reported that the RF method yielded 100% classification accuracy in detecting congestive heart failure when compared with other classifiers. RF is a stochastic ensemble learning algorithm. With the combination of multiple decision trees at different subsets of data set, RF can effectively handle multitudinous high-dimensional data, and meanwhile avoid over-fitting [ 34 ]. Depending on the fast training speed, RF can quickly sort the importance of variables, and achieve the highest accuracy [ 32 – 34 ]. Our results further corroborates the robustness of RF in the application of predicting treatment response. It is well-known that clinical and pathological factors are essential predictors for treatment effect [ 2 , 22 ]. Since the introduction of HER2-targeted drugs, such as trastuzumab and pertuzumab, HER2 + has been regarded as a good prognostic factors of treatment response [ 35 ]. In contrast, the presence of hormone receptor (HR) positive has been reported as a poor prognostic factors of NAC [ 22 ]. In agreement with previous studies, our results showed that pCR to NAC occurred most commonly in HER2 + patients, and the HR+/HER2- patients were the least sensitive to complete response [ 21 , 35 ]. Previous studies also indicated that tumor size, tumor grade, and Ki67 levels were predictors associated with treatment effect [ 36 ]. However, in the present study, tumor size, clinical stage, histologic type, and Ki67 levels all had no significant difference between the pCR and non-pCR groups. The results suggested that the molecular subtypes of tumor may be more important than tumor size, stage, and proliferation in the determination of pCR status. The strengths of this study contributes to the fields of breast US radiomics in the following ways: First, it is the first attempt to explore the role of peritumoral environment based on US radiomics, in comparison with intratumoral region and clinicopathologic factors to predict pCR after NAC in LABC patients. Second, we applied five popular ML algorithms as a new approach to construct various predictive models, which may provide more convincing results. Third, compared to prior radiomics studies, which mainly focused on MRI or other image-based radiomics for monitoring NAC, US is a more convenient diagnostic tool in breast examination, thus making the US-based radiomics analysis with more wide application in clinical practice. Therefore, the present study represents a possible novel confluence of US radiomics to estimate peritumoral response, and is a further step toward “perfect” predictive tool to guide more individual treatment strategy. Certainly, our study has some limitations: First, it was a retrospective study performed in a single institution. Second, we only included mass lesions with sufficient peritumoral tissue identified on US image, nonmass-like lesions and lesions with no clear peritumoral region on US images were excluded which may cause selected bias. In addition, we didn’t construct molecular subtypes models to assess the predictive performance because of the relative small number of subjects in each subgroup. Finally, our study did not include genomics data, but that was beyond the scope of our manuscript, further studies would be expected to address this issue in the future. In conclusion, although the RF classifier-based PURS yielded relatively lower efficacy than IURS model for the prediction of pCR, it also achieved a favorable predictive accuracy, with AUCs higher than clinicopathologic predictors. The consideration of PURS with ML algorithm as a novel and promising predictive tool may aid in clinical systemic therapy decision-making for LABC patients receiving NAC. Abbreviations AdaBoost : Adaptive boosting, AUC: Area under the curve, BI-RADS: Breast Imaging and Reporting Data System, CI: Confidence interval, ER: Estrogen receptor, GLCM: Gray-level co-occurrence matrix, GLDM: Gray-level dependence matrix, GLRLM: Gray-level run-length matrix, GLSZM: Gray-level size-zone matrix, HER2: Human epidermal growth factor receptor 2, ICC: Intra-class correlation coefficient, IURS: Intratumoral ultrasound radiomics signatures, LABC: locally advanced breast cancer, LDA: Linear discriminant analysis, LR: logistic regression, NAC: Neoadjuvant chemotherapy, NGTDM: Neighboring gray tone difference matrix, PCR: Pathological complete response, PURS: Peritumoral ultrasound radiomics signatures, RF: Random forest, RFE: Recursive feature elimination, PR: Progesterone receptor, ROC: Receiver operating characteristic curve, SD: Standard deviation, SVM: Support vector machine, US: Ultrasound Declarations Author Contribution Authors' contributions were as following: Jiejie Yao, Wei Zhou and Jianqiao Zhou had primary responsibility for the protocol development, patient enrolment, preliminary data analysis, and writing of the draft. Xiaohong Jia and Ying Zhu analyzed the data. Xiaosong Chen and Weiwei Zhan assisted with the data collection and verification. Jianqiao Zhou supervised data collection, and reviewed the manuscript for important intellectual content. Wei Zhou and Jianqiao Zhou supervised the design and execution of the study, contributed to the writing of the manuscript and had final approval of the manuscript submitted.All authors confirmed that they had full access to all the data in the study and accept responsibility to submit for publication. References Korde LA, Somerfield MR, Carey LA, Crews JR, Denduluri N, Hwang ES, et al (2021) Neoadjuvant chemotherapy, endocrine therapy, and targeted therapy for breast cancer: ASCO Guideline. J Clin Oncol 39:1485-1505. doi: 10.1200/JCO.20.03399. Wang H, Mao X (2020) Evaluation of the efficacy of neoadjuvant chemotherapy for breast cancer. Drug Des Devel Ther.14:2423-2433. doi: 10.2147/DDDT.S253961. Chen S, Chen CM, Yu KD, Yang WT, Shao ZM (2012) A prognostic model to predict outcome of patients failing to achieve pathological complete response after anthracycline-containing neoadjuvant chemotherapy for breast cancer. J Surg Oncol. 105:577-85. doi: 10.1002/jso.22140. M Braden A, V Stankowski R, M Engel J, A Onitilo A (2014) Breast cancer biomarkers: risk assessment, diagnosis, prognosis, prediction of treatment efficacy and toxicity, and recurrence. Curr Pharm Des. 20:4879-98. doi: 10.2174/1381612819666131125145517. Liu Z, Li Z, Qu J, Zhang R, Zhou X, Li L, et al (2019) Radiomics of multiparametric MRI for pretreatment prediction of pathologic complete response to neoadjuvant chemotherapy in breast cancer: a multicenter study. Clin Cancer Res 25:3538-3547. doi: 10.1158/1078-0432.CCR-18-3190. Bian T, Wu Z, Lin Q, Wang H, Ge Y, Duan S, et al (2020) Radiomic signatures derived from multiparametric MRI for the pretreatment prediction of response to neoadjuvant chemotherapy in breast cancer. Br J Radiol. 93:20200287. doi: 10.1259/bjr.20200287. Huang X, Mai J, Huang Y, He L, Chen X, Wu X, et al. Radiomic nomogram for pretreatment prediction of pathologic complete response to neoadjuvant therapy in breast cancer: predictive value of staging contrast-enhanced CT. Clin Breast Cancer 2021;21:e388-e401. doi: 10.1016/j.clbc.2020.12.004. Lee H, Lee DE, Park S, Kim TS, Jung SY, Lee S, et al (2019) Predicting response to neoadjuvant chemotherapy in patients with breast cancer: combined statistical modeling using clinicopathological factors and FDG PET/CT texture parameters. Clin Nucl Med. 44:21-29. doi: 10.1097/RLU.0000000000002348. Zhang MQ, Du Y, Zha HL, Liu XP, Cai MJ, Chen ZH, et al (2022) Construction and validation of a personalized nomogram of ultrasound for pretreatment prediction of breast cancer patients sensitive to neoadjuvant chemotherapy. Br J Radiol. 95:20220626. doi: 10.1259/bjr.20220626. Braman NM, Etesami M, Prasanna P, Dubchuk C, Gilmore H, Tiwari P, et al (2017) Intratumoral and peritumoral radiomics for the pretreatment prediction of pathological complete response to neoadjuvant chemotherapy based on breast DCE-MRI. Breast Cancer Res.19:57. doi: 10.1186/s13058-017-0846-1. Mao N, Shi Y, Lian C, Wang Z, Zhang K, Xie H, et al (2022) Intratumoral and peritumoral radiomics for preoperative prediction of neoadjuvant chemotherapy effect in breast cancer based on contrast-enhanced spectral mammography. Eur Radiol 32:3207-3219. doi: 10.1007/s00330-021-08414-7. Haug CJ, Drazen JM (2023) Artificial intelligence and machine learning in clinical medicine, 2023. N Engl J Med 388:1201-1208. doi.org/10.1056/NEJMra2302038. Yao J, Zhou W, Xu S, Jia X, Zhou J, Chen X, et al (2023) Machine learning-based breast tumor ultrasound radiomics for pre-operative prediction of axillary sentinel lymph node metastasis burden in early-stage invasive breast cancer. Ultrasound Med Biol. 10:S0301-5629(23)00335-6. doi: 10.1016/j.ultrasmedbio.2023.10.004. Penault-Llorca F, André F, Sagan C, Lacroix-Triki M, Denoux Y, Verriele V, et al (2009) Ki67 expression and docetaxel efficacy in patients with estrogen receptor-positive breast cancer. J Clin Oncol 27:2809-2815. doi: 10.1200/JCO.2008.18.2808. Gradishar WJ, Moran MS, Abraham J, Aft R, Agnese D, Allison KH, et al (2022) Breast Cancer, Version 3.2022, NCCN Clinical Practice Guidelines in Oncology. J Natl Compr Canc Netw 20:691-722. doi: 10.6004/jnccn.2022.0030. Jones RL, Lakhani SR, Ring AE, Ashley S, Walsh G, Smith IE (2006) Pathological complete response and residual DCIS following neoadjuvant chemotherapy for breast carcinoma. Br J Cancer 94:358-362. doi: 10.1038/sj.bjc.6602950. Mercado CL (2014) BI-RADS update. Radiol Clin North Am 52:481-487. doi: 10.1016/j.rcl.2014.02.008. Fedorov A, Beichel R, Kalpathy-Cramer J, Finet J, Fillion-Robin JC, Pujol S, et al (2012) 3D Slicer as an image computing platform for the quantitative imaging network. Magn Reson Imaging 30:1323-1341. doi: 10.1016/j.mri.2012.05.001. Xu Z, Shen D, Kou Y, Nie T (2022) A synthetic minority oversampling technique based on gaussian mixture model filtering for imbalanced data classification. IEEE Trans Neural Netw Learn Syst. PP. doi: 10.1109/TNNLS.2022.3197156. Kernbach JM, Staartjes VE (2022) Foundations of machine learning-based clinical prediction modeling: Part II-generalization and overfitting. Acta Neurochir Suppl 134:15-21. doi:10.1007/978-3-030-85292-4_3. Song Y, Zhang J, Zhang YD, Hou Y, Yan X, Wang Y, et al (2020) FeAture explorer (FAE): A tool for developing and comparing radiomics models. PLoS One 15:e0237587. doi:10.1371/journal.pone.0237587. von Minckwitz G, Untch M, Blohmer JU, Costa SD, Eidtmann H, Fasching PA, et al (2012) Definition and impact of pathologic complete response on prognosis after neoadjuvant chemotherapy in various intrinsic breast cancer subtypes. J Clin Oncol. 30:1796-1804. doi: 10.1200/JCO.2011.38.8595. Savaridas SL, Sim YT, Vinnicombe SJ, Purdie CA, Thompson AM, Evans A (2019) Are baseline ultrasound and mammographic features associated with rates of pathological completes response in patients receiving neoadjuvant chemotherapy for breast cancer? Cancer Imaging 19:67. doi: 10.1186/s40644-019-0251-3. Jiang M, Li CL, Luo XM, Chuan ZR, Lv WZ, Li X, et al (2012) Ultrasound-based deep learning radiomics in the assessment of pathological complete response to neoadjuvant chemotherapy in locally advanced breast cancer. Eur J Cancer 147:95-105. doi: 10.1016/j.ejca.2021.01.028. Cheon H, Kim HJ, Kim TH, Ryeom HK, Lee J, Kim GC, et al (2018) Peritumoral invasive breast cancer: prognostic value of peritumoral edema identified at preoperative MR imaging. Radiology 287:68-75. doi: 10.1148/radiol.2017171157. Li J, Guo L, Yin L, Fang H, Ye W, Zhao B, et al (2018) Can different regions of interest influence the diagnosis of benign and malignant breast lesions using quantitative parameters of contrast-enhanced sonography? Eur J Radiol. 108:1-6. doi: 10.1016/j.ejrad.2018.09.005. Abbasian Ardakani A, Bureau NJ, Ciaccio EJ, Acharya UR (2022) Interpretation of radiomics features - a pictorial review. Comput Methods Programs Biomed 215:106609. doi: 10.1016/j.cmpb.2021.106609. van Griethuysen JJM, Fedorov A, Parmar C, Hosny A, Aucoin N, Narayan V, et al (2017) Computational radiomics system to decode the radiographic phenotype. Cancer Res 77:e104-e107. doi: 10.1158/0008-5472. Sudarshan VK, Mookiah MR, Acharya UR, Chandran V, Molinari F, Fujita H, et al (2016) Application of wavelet techniques for cancer diagnosis using ultrasound images: a review. Comput Biol Med. 69:97-111. doi: 10.1016/j.compbiomed.2015.12.006. Tang B, Chen Y, Wang Y, Nie J (2021) A wavelet-based learning model enhances molecular prognosis in pancreatic adenocarcinoma. Biomed Res Int. 7865856. doi: 10.1155/2021/7865856. Tahmassebi A, Wengert GJ, Helbich TH, Bago-Horvath Z, Alaei S, Bartsch R, et al (2019) Impact of machine learning with multiparametric magnetic resonance imaging of the breast for early prediction of response to neoadjuvant chemotherapy and survival outcomes in breast cancer patients. Invest Radiol. 54:110-117. doi: 10.1097/RLI.0000000000000518. Sun K, Zhu H, Chai WM, Yan FH (2023) TP53 mutation estimation based on MRI radiomics analysis for breast cancer. J Magn Reson Imaging 57:1095-1103. doi: 10.1002/jmri.28323. Masetic Z, Subasi A (2016) Congestive heart failure detection using random forest classifier. Comput Methods Programs Biomed. 130:54-64. doi: 10.1016/j.cmpb.2016.03.020. Rigatti SJ (2017) Random Forest. J Insur Med. 47:31-39. doi: 10.17849/insm-47-01-31-39.1. Takada M, Toi M (2020) Neoadjuvant treatment for HER2-positive breast cancer. Chin Clin Oncol. 9:32. doi: 10.21037/cco-20-123. Keam B, Im SA, Park S, Nam BH, Han SW, Oh DY, et al (2011) Nomogram predicting clinical outcomes in breast cancer patients treated with neoadjuvant chemotherapy. J Cancer Res Clin Oncol. 137:1301-1308. doi: 10.1007/s00432-011-0991-3. Additional Declarations No competing interests reported. Supplementary Files AppendixADetailsofUSradiomicsfeatures.docx AppendixBDetailsofthe851extractedradiomicsfeatures.xlsx AppendixCGTheselectedfeaturesofMLbasedPURSmodels.docx AppendixHLTheselectedfeaturesinMLbasedIURSmodels.docx Cite Share Download PDF Status: Posted Version 1 posted 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 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-4440501","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":307154864,"identity":"bdbb41e3-9ce8-483c-88b7-9bdc499daf15","order_by":0,"name":"Jiejie Yao","email":"","orcid":"","institution":"Ruijin Hospital, Shanghai Jiao Tong University School of Medicine","correspondingAuthor":false,"submittingAuthor":false,"prefix":"","firstName":"Jiejie","middleName":"","lastName":"Yao","suffix":""},{"id":307154869,"identity":"019e0597-e9f7-4d00-b717-72eefafb7723","order_by":1,"name":"Wei Zhou","email":"","orcid":"","institution":"Ruijin Hospital, Shanghai Jiao Tong University School of Medicine","correspondingAuthor":false,"submittingAuthor":false,"prefix":"","firstName":"Wei","middleName":"","lastName":"Zhou","suffix":""},{"id":307154873,"identity":"0a344857-c390-4fd3-9916-be79cd7180ac","order_by":2,"name":"Xiaohong Jia","email":"","orcid":"","institution":"Ruijin Hospital, Shanghai Jiao Tong University School of Medicine","correspondingAuthor":false,"submittingAuthor":false,"prefix":"","firstName":"Xiaohong","middleName":"","lastName":"Jia","suffix":""},{"id":307154875,"identity":"d9e4d719-6886-4efd-b733-6ebb6c8a1ff0","order_by":3,"name":"Ying Zhu","email":"","orcid":"","institution":"Ruijin Hospital, Shanghai Jiao Tong University School of Medicine","correspondingAuthor":false,"submittingAuthor":false,"prefix":"","firstName":"Ying","middleName":"","lastName":"Zhu","suffix":""},{"id":307154876,"identity":"fbe50e4d-a2f8-4020-b8f2-3b857e41d6a2","order_by":4,"name":"Xiaosong Chen","email":"","orcid":"","institution":"Ruijin Hospital, Shanghai Jiao Tong University School of Medicine","correspondingAuthor":false,"submittingAuthor":false,"prefix":"","firstName":"Xiaosong","middleName":"","lastName":"Chen","suffix":""},{"id":307154878,"identity":"bc7daced-612d-4a41-a89c-82c1cac99799","order_by":5,"name":"Weiwei Zhan","email":"","orcid":"","institution":"Ruijin Hospital, Shanghai Jiao Tong University School of Medicine","correspondingAuthor":false,"submittingAuthor":false,"prefix":"","firstName":"Weiwei","middleName":"","lastName":"Zhan","suffix":""},{"id":307154881,"identity":"e18329d0-21b7-4897-b123-db2203748480","order_by":6,"name":"Jianqiao Zhou","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAAu0lEQVRIiWNgGAWjYDACCTBZw8PPzHzwASlajslJtrMlG5CihdnY4DyPmQBROuRnNx+T/NnGlrj5MIMZ0IE20QS1MM45libN2yaTuO0wQ9oDhmNpuQ2EtDBL5JhJMwJtAWo5bsDYcJiwFjagFqDDmBM3NzO2SRClhQeoRYK3Deh9ZmY24rRISKQlW/OcOyYncZiN2SCBGL/Iz0g+ePNHGTAq+89/fPChxoawFlSQQJryUTAKRsEoGAW4AAB7RTdpKnCypQAAAABJRU5ErkJggg==","orcid":"","institution":"Ruijin Hospital, Shanghai Jiao Tong University School of Medicine","correspondingAuthor":true,"submittingAuthor":false,"prefix":"","firstName":"Jianqiao","middleName":"","lastName":"Zhou","suffix":""}],"badges":[],"createdAt":"2024-05-18 09:55:22","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-4440501/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-4440501/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":57948151,"identity":"17bb9d38-df1d-475c-a5ff-70ce14d2fda0","added_by":"auto","created_at":"2024-06-07 20:31:34","extension":"jpg","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":590095,"visible":true,"origin":"","legend":"\u003cp\u003eFlowchart shows patient recruitment and study design. A total of 505 LABC patients with pretreatment breast tumor US images, and accepted NAC and post NAC surgery at our institution were firstly included. Among them, 358 patients finally met the criteria and divided into a training set and a test set according to the examination time.\u003c/p\u003e","description":"","filename":"fig.1.jpg","url":"https://assets-eu.researchsquare.com/files/rs-4440501/v1/6c03dd9d87c941c7ab4671d2.jpg"},{"id":57948149,"identity":"084407c6-2b3b-443d-b22e-e8f2091f9d52","added_by":"auto","created_at":"2024-06-07 20:31:34","extension":"jpg","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":1174721,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003e(a-h) \u003c/strong\u003eExamples of the regions of interest (ROIs) segmentation. \u003cstrong\u003e(a) \u003c/strong\u003eThe baseline grayscale US image of a 48-year-old woman with primary left breast tumor who attained pCR after NAC, \u003cstrong\u003e(e) \u003c/strong\u003ethe baseline grayscale US image of a 60-year-old woman with primary right breast tumor who yielded non-pCR after NAC, \u003cstrong\u003e(b, f)\u003c/strong\u003e the corresponding ROIs were manually delineated along the contour of the tumor, \u003cstrong\u003e(c, g)\u003c/strong\u003e the ROIs of peritumoral regions (yellow), and \u003cstrong\u003e(d, h) \u003c/strong\u003ethe ROIs of intratumoral regions (green).\u003c/p\u003e","description":"","filename":"Figure2.jpg","url":"https://assets-eu.researchsquare.com/files/rs-4440501/v1/9aca23ee2402f23e1fe3af82.jpg"},{"id":57948152,"identity":"ca12d49c-921f-4f96-88af-081c5a7cbc72","added_by":"auto","created_at":"2024-06-07 20:31:35","extension":"jpg","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":1090530,"visible":true,"origin":"","legend":"\u003cp\u003eThe overview of the workflow. \u003cstrong\u003e1.\u003c/strong\u003e Baseline grayscale US images of 358 breast tumors and their corresponding peritumoral and intratumoral ROIs. \u003cstrong\u003e2.\u003c/strong\u003e Extraction of radiomic features including shape features, first-order features, texture features (i.e., GLCM, GLDM, GLRLM, GLSZM, and NGTDM), and wavelet-related features. \u003cstrong\u003e3.\u003c/strong\u003e SMOTE, Z-score, Mean normalization, PCA, PCC, and RFE were employed in the features selection. \u003cstrong\u003e4.\u003c/strong\u003e Five ML algorithms (i.e., LDA, SVM, RF, LR, and Adaboost classifiers) were applied to construct PURS and IURS radiomics models for the prediction of pCR to NAC. \u003cstrong\u003e5.\u003c/strong\u003e The predictive efficacy of models.\u003c/p\u003e","description":"","filename":"Figure3.jpg","url":"https://assets-eu.researchsquare.com/files/rs-4440501/v1/477fbd249fcb8c0d78f618b3.jpg"},{"id":57948153,"identity":"77366964-af38-4de8-b3be-f8d35aed2a31","added_by":"auto","created_at":"2024-06-07 20:31:35","extension":"jpg","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":712260,"visible":true,"origin":"","legend":"\u003cp\u003eThe ROC curves of PURS model with RF \u003cstrong\u003e(a)\u003c/strong\u003e, LR \u003cstrong\u003e(b)\u003c/strong\u003e, Adaboost \u003cstrong\u003e(c)\u003c/strong\u003e, SVM \u003cstrong\u003e(d)\u003c/strong\u003e, and LDA \u003cstrong\u003e(e) \u003c/strong\u003eclassifiers. The AUCs for the five classifiers were (0.882, 0.866, 0.834, 0.738, and 0.687) in the training set, and (0.889, 0.849, 0.823, 0.746, and 0.732) in the test set, respectively.\u003c/p\u003e","description":"","filename":"Figure4.jpg","url":"https://assets-eu.researchsquare.com/files/rs-4440501/v1/ef85aa3289f847ee50eefaae.jpg"},{"id":57948960,"identity":"21258cfc-6d42-4b82-a8b3-94ecb95971e6","added_by":"auto","created_at":"2024-06-07 20:39:34","extension":"jpg","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":701512,"visible":true,"origin":"","legend":"\u003cp\u003eThe ROC curves of IURS model with RF \u003cstrong\u003e(a)\u003c/strong\u003e, Adaboost \u003cstrong\u003e(b)\u003c/strong\u003e,\u003cstrong\u003e \u003c/strong\u003eLR \u003cstrong\u003e(c)\u003c/strong\u003e, SVM \u003cstrong\u003e(d)\u003c/strong\u003e, and LDA \u003cstrong\u003e(e) \u003c/strong\u003eclassifiers. The AUCs for the five classifiers were (0.948, 0.910, 0.852, 0.817, and 0.825) in the training set, and (0.931, 0.920, 0.875, 0.825, and 0.798) in the test set, respectively.\u003c/p\u003e","description":"","filename":"Figure5.jpg","url":"https://assets-eu.researchsquare.com/files/rs-4440501/v1/25271fd920aa5dc12d677548.jpg"},{"id":57948158,"identity":"5ec21c83-f254-4d67-a3c3-4620f411ba8e","added_by":"auto","created_at":"2024-06-07 20:31:36","extension":"jpg","order_by":6,"title":"Figure 6","display":"","copyAsset":false,"role":"figure","size":554502,"visible":true,"origin":"","legend":"\u003cp\u003eThe ROC curves of RF-based IURS, PURS and clinicopathologic models in the training \u003cstrong\u003e(a)\u003c/strong\u003e and test \u003cstrong\u003e(b)\u003c/strong\u003e sets. The AUCs for the RF-based IURS model (yellow curves), PURS model (purple curves), and C model (blue curves) in both sets were (0.948, 0.882, and 0.776) and ( 0.931, 0.889, and 0.759), respectively.\u003c/p\u003e","description":"","filename":"Figure6.jpg","url":"https://assets-eu.researchsquare.com/files/rs-4440501/v1/ef382db77f1cc0f271880ad6.jpg"},{"id":57949608,"identity":"66415f12-3fdb-432e-baff-e3e911790067","added_by":"auto","created_at":"2024-06-07 20:47:40","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":5712896,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-4440501/v1/99cacbaa-1a5e-4a6b-bd15-4552ee03251f.pdf"},{"id":57948155,"identity":"6e4c79f6-e984-46d6-84a3-ad352b4c3656","added_by":"auto","created_at":"2024-06-07 20:31:35","extension":"docx","order_by":9,"title":"","display":"","copyAsset":false,"role":"supplement","size":12997,"visible":true,"origin":"","legend":"","description":"","filename":"AppendixADetailsofUSradiomicsfeatures.docx","url":"https://assets-eu.researchsquare.com/files/rs-4440501/v1/e87228994c3ff4d66f57a50a.docx"},{"id":57948961,"identity":"37810e3a-fbe6-4eec-a51a-140222f1b0e4","added_by":"auto","created_at":"2024-06-07 20:39:35","extension":"xlsx","order_by":10,"title":"","display":"","copyAsset":false,"role":"supplement","size":33216,"visible":true,"origin":"","legend":"","description":"","filename":"AppendixBDetailsofthe851extractedradiomicsfeatures.xlsx","url":"https://assets-eu.researchsquare.com/files/rs-4440501/v1/754259bc6d08cad8b0a53ea1.xlsx"},{"id":57948157,"identity":"aefc70b9-302f-4b09-9b0d-1af9196d09fb","added_by":"auto","created_at":"2024-06-07 20:31:35","extension":"docx","order_by":11,"title":"","display":"","copyAsset":false,"role":"supplement","size":15881,"visible":true,"origin":"","legend":"","description":"","filename":"AppendixCGTheselectedfeaturesofMLbasedPURSmodels.docx","url":"https://assets-eu.researchsquare.com/files/rs-4440501/v1/62129e8914cfe78b534e5b87.docx"},{"id":57948159,"identity":"bdef2a0d-cea9-4bca-81ec-13ee8d5b9333","added_by":"auto","created_at":"2024-06-07 20:31:36","extension":"docx","order_by":12,"title":"","display":"","copyAsset":false,"role":"supplement","size":16362,"visible":true,"origin":"","legend":"","description":"","filename":"AppendixHLTheselectedfeaturesinMLbasedIURSmodels.docx","url":"https://assets-eu.researchsquare.com/files/rs-4440501/v1/2d223e4fb1aac5ab5fa1baa6.docx"}],"financialInterests":"No competing interests reported.","formattedTitle":"Machine learning prediction of pathological complete response to neoadjuvant chemotherapy with peritumoral breast tumor ultrasound radiomics: compare with intratumoral radiomics and clinicopathologic predictors","fulltext":[{"header":"Introduction","content":"\u003cp\u003eFor patients with locally advanced breast cancer (LABC), neoadjuvant chemotherapy (NAC) has been a standard treatment strategy to downstage tumor, reduce metastasis, and improve the probability of breast-conserving surgery [\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e]. Ideally, it could infer a favorable disease-free and an improved overall survival when a pathological complete response (pCR) was achieved after NAC [\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e]. However, according to the previous reports, approximately 30% of the patients may not respond to NAC, or even experience disease progression due to the morphological heterogeneous of tumors, or the genetic mutation [\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e, \u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e]. Genetic identification is expensive and not routinely performed in clinical practice. Histopathologic examination is notable for its invasive procedure, and inevitable complications such as pain, hematoma and infection. Also, for heterogeneous tumor with different responses to NAC, a small tissue provided by biopsy is apt to selection bias [\u003cspan additionalcitationids=\"CR3\" citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e]. Noninvasive, accurate and novel approaches to predict patients who will achieve pCR, and those who will benefit less but suffer more toxicities from NAC are highly desired.\u003c/p\u003e \u003cp\u003eRadiomics can extract a large number of high-dimensional data from traditional medical images, and has been applied to predict therapeutic response. However, most of previous radiomics studies based on MRI [\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e, \u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e], or CT [\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e], or PET/CT images [\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e], only few studies utilized US radiomics for prediction [\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e]. Moreover, prior researchers mainly focused on parameters within intratumoral structure, while the considerations about peritumoral region, which has been described as a \u0026ldquo;reactive zone\u0026rdquo; surrounding the tumor were limited. Recently, researchers began to explore intratumoral\u0026ensp;and\u0026ensp;peritumoral\u0026ensp;radiomics\u0026ensp;to predict NAC effect based on MRI [\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e] or mammography [\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e]. As a convenient, low cost, and widely used diagnostic tool, US plays an important role in screening and evaluating tumor margins. However, to our knowledge, no prior study used peritumoral US radiomics signatures (PURS) for the prediction of pCR to NAC.\u003c/p\u003e \u003cp\u003eCompared to previous radiomics studies, most of which used the least absolute shrinkage and selection operator (LASSO) regression to establish predictive nomogram [\u003cspan additionalcitationids=\"CR6 CR7 CR8\" citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e], machine learning (ML) algorithms such as linear discriminant analysis (LDA), support vector machine (SVM), random forest (RF), logistic regression (LR) and adaptive boosting (Adaboost) can handle abundant quantitative radiomics features powerfully and effectively [\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e]. Nevertheless, the application of ML-based PURS and IURS to predict NAC effect has not been explored. A prior study of our team has constructed four ML classifiers-based US radiomics for preoperative prediction of axillary sentinel lymph node metastasis burden in early-stage BC patients [\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e]. Now, we were interested in whether PURS can be employed to predict pCR after NAC. We also wondered the predictive ability of various ML algorithms based radiomics models.\u003c/p\u003e \u003cp\u003eTherefore, the purpose of this study was to assess the efficacy of five ML classifiers-based PURS, compared with IURS and clinicopathologic factors, for early prediction of pCR after NAC in LABC patients.\u003c/p\u003e"},{"header":"Methods","content":"\u003cdiv id=\"Sec3\" class=\"Section2\"\u003e \u003ch2\u003eStudy population\u003c/h2\u003e \u003cp\u003eThis study was a retrospective analysis, and was conducted in accordance with the Declaration of Helsinki, approved by the Ethics Committee of our institution. The informed consent was waived due to its retrospective nature. Data sets were obtained from January 2018 to December 2022, we firstly enrolled 505 LABC patients who underwent pretreatment breast tumor biopsy, then accepted NAC and post NAC surgery at our institution. The inclusion criteria were as follows: (i) patients with biopsy-proven primary BC, and without distant metastasis; (ii) patients with high quality baseline breast tumor US images before biopsy; (iii) patients underwent a full course of NAC; (iv) patients accepted surgery after NAC, and pCR or non-pCR status was confirmed by surgical specimen histopathologic examinations. The exclusion criteria were as follows: (i) patients who did not complete NAC regimen; (ii) patients with bilateral BC or multiple tumors; (iii) patients with nonmass-like lesions; (iv) patients with no sufficient peritumoral tissue identified on US images. Finally, a total of 358 patients (all women, mean age, 48.1 years\u0026plusmn;\u0026thinsp;10.5; median age, 49.6 years; age range, 38\u0026ndash;79 years) were included, and were divided into a training set (from January 2018 to December 2020, n\u0026thinsp;=\u0026thinsp;250), and an independent test set (from January 2021 to December 2022, n\u0026thinsp;=\u0026thinsp;108). Figure\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e shows the patient recruitment and study design.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec4\" class=\"Section2\"\u003e \u003ch2\u003eData collection\u003c/h2\u003e \u003cp\u003eThe clinical data such as patients\u0026rsquo; age, clinical T and N stage, and NAC regimens were recorded and retrieved from the Shanghai Jiaotong University Breast Cancer Database (SJTU-BCDB). The pathological data included tumor histological type, tumor proliferation rate (Ki67 levels), estrogen receptor (ER), progesterone receptor (PR), and human epidermal growth factor receptor 2 (HER2) status. HER2 status was confirmed with fluorescence in situ hybridization. A cut-off value for Ki67 positive was established at 20% [\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e]. Tumors were classified into 3 subgroups based on the expression of ER, PR, and HER2 status: HR+/HER2- (HR+, HER2-); HER2+ (HER2+, ER\u0026thinsp;+\u0026thinsp;or ER-, PR\u0026thinsp;+\u0026thinsp;or PR-); and triple-negative (ER-, PR-, HER2-). HR\u0026thinsp;+\u0026thinsp;was defined as ER\u0026thinsp;+\u0026thinsp;and/or PR+. All the patients received six or eight cycles of NAC before breast surgery according to the National Comprehensive Cancer Network (NCCN) guideline [\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e]. The pCR is defined as no residual invasive (ductal carcinoma in situ could be present), and no axillary lymph node invasion in the final post NAC surgical specimen (ypT0/isN0) [\u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e].\u003c/p\u003e \u003cdiv id=\"Sec5\" class=\"Section3\"\u003e \u003ch2\u003eUS images acquisition and segmentation\u003c/h2\u003e \u003cp\u003eBreast US examinations were performed 1 week before biopsy by Resona 7 (Mindray Medical International, Shenzhen, China) with a linear probe at 3\u0026ndash;11 MHz, and Esaote MyLab 60 (Esaote, Genoa, Italy) with a linear probe at 4\u0026ndash;13 MHz. Tumors were assessed according to the Breast Imaging Reporting and Data System (BI-RADS) [\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e]. The maximum size of the breast tumors measured by US were also recorded.\u003c/p\u003e \u003cp\u003eThe tumor regions of interest (ROIs) segmentation were performed via a free open-source software package (3D Slicer version 5.0.3) [\u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e]. The intratumoral region was segmented by dilating the delineated tumor contour manually in the largest cross-sectional area. The peritumoral ROI was obtained with a 3mm-thick surrounding zone outside the intratumoral region automatically using \u0026ldquo;Hollow\u0026rdquo; and \u0026ldquo;Margin\u0026rdquo; segment editors. The ROIs of peritumoral and intratumoral area were extracted separately with PyRadiomics software [\u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e]. Figure\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e (a-h) shows the examples of tumor US images and its corresponding ROIs.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eIntra- and interobserver reproducibility were assessed with intra- and interclass correlation coefficients (ICCs). Two experienced radiologists (author 1 and author 2, with over fifteen years of experience in breast US, and three years of experience in the software) who were blinded to the treatment outcomes, segmented the peritumoral and intratumoral regions of 60 randomly selected breast tumors, and extracted the radiomics signatures separately. One weeks later, author 1 repeated the same procedure and analyzed the remaining images. An ICC equal to or higher than 0.75 was considered as good intra- and interobserver agreement, and was included in the further feature selection process.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec6\" class=\"Section3\"\u003e \u003ch2\u003eFeature extraction, selection and ML-classifiers implementation\u003c/h2\u003e \u003cp\u003eFor each breast tumor, a total of 851 radiomics features were extracted and categorized as follows: 14 shape features, 18 first-order features, 24 gray level co-occurrence matrix (GLCM), 14 gray level dependence matrix (GLDM), 16 gray level run length matrix (GLRLM), 16 gray level size zone matrix (GLSZM), 5 neighboring gray tone difference matrix (NGTDM), and 744 wavelet-related features (details are shown in Supplementary Appendix A and B). The final training sets comprised each of 212,750 PURS and IURS, the independent test sets contained each of 91,908 radiomics signatures.\u003c/p\u003e \u003cp\u003eTo select the most robust features, synthetic minority oversampling technique (SMOTE) was firstly used to remove the unbalance samples in the data set [\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e]. Then, Z-score and Mean normalization methods were applied to standardize the corresponding features. Principal component analysis (PCA) and pearson correlation coefficient (PCC) were employed to increase data interpretation and reduce feature dimension. After that, recursive features elimination (RFE) was used to detect the most relevant signatures [\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e]. Finally, the most significant selected features were input to LDA, SVM, RF, LR, and Adaboost classifiers with a 5-fold cross validation to construct PURS and IURS models for the prediction of pCR. Figure\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003e shows the overview of the workflow.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv id=\"Sec7\" class=\"Section2\"\u003e \u003ch2\u003eStatistical Analysis\u003c/h2\u003e \u003cp\u003eAll numerical data were presented as mean\u0026thinsp;\u0026plusmn;\u0026thinsp;standard deviation. Continuous and categorical variables were compared using the independent \u003cem\u003et\u003c/em\u003e test and the Chi-square test or Fisher\u0026rsquo;s exact test, respectively. All clinical and pathological factors that were shown to be potentially associated with pCR (\u003cem\u003ep\u0026thinsp;\u0026lt;\u003c/em\u003e\u0026thinsp;0.05 in univariate analysis) were considered to construct a clinicopathologic model. The training data set was used to construct five ML classifiers-based PURS and IURS predictive models, the test data set was used for independent validation to evaluate the performance of the models. The predictive abilities were assessed with respect to sensitivity (SEN), specificity (SPE), accuracy (ACC), positive predictive value (PPV), negative predictive value (NPV), and the area under the receiver operating characteristic (ROC) curve (AUC). Comparisons between AUCs were made by using the DeLong test. All of the processes were implemented with FeAture Explorer Pro (FAEPro, V0.5.3) in Python (3.7.6) [\u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e] \u003cem\u003e(\u003c/em\u003e\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://github.com/salan668/FAE\u003c/span\u003e\u003cspan address=\"https://github.com/salan668/FAE\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003cem\u003e)\u003c/em\u003e, SPSS software (version 23.0), and MedCalc software (version 22.013). A \u003cem\u003eP\u003c/em\u003e value less than 0.05 was regarded as statistically significant difference.\u003c/p\u003e \u003c/div\u003e"},{"header":"Results","content":"\u003cdiv id=\"Sec9\" class=\"Section2\"\u003e\n \u003ch2\u003eClinicopathologic characteristics\u003c/h2\u003e\n \u003cp\u003eAmong all the 358 patients, the rate of pCR was 27.1% (97/358). It was 27.2% (68/250) in the training set, and 26.9% (29/108) in the test set. The ER, PR, and HER2 status were found to be significantly associated with pCR in both sets (\u003cem\u003ep\u003c/em\u003e\u0026thinsp;\u0026lt;\u0026thinsp;0.05 for all). Meanwhile, no significant difference was detected in terms of age, tumor maximum size, clinical T and N stage, histologic type and Ki-67 levels between pCR and non-pCR groups (\u003cem\u003ep\u0026thinsp;\u0026gt;\u003c/em\u003e\u0026thinsp;0.05 for all). Regarding molecular subtypes, the rate of pCR was significantly higher in HER2\u0026thinsp;+\u0026thinsp;patients [58.8% (40/68), 69.0% (20/29)], than triple-negative [25.0% (17/68), 17.2% (5/29)], and HR+/HER2- patients [16.2% (11/68), 13.8% (4/29)] in the training and test sets, respectively (\u003cem\u003ep\u003c/em\u003e\u0026thinsp;\u0026lt;\u0026thinsp;0.001 for all) (Table \u003cspan class=\"InternalRef\"\u003e1\u003c/span\u003e).\u003c/p\u003e\n \u003cdiv class=\"gridtable\"\u003e\u0026nbsp;\u003ctable id=\"Tab1\" border=\"1\"\u003e\n \u003ccaption language=\"En\"\u003e\n \u003cdiv class=\"CaptionNumber\"\u003eTable 1\u003c/div\u003e\n \u003cdiv class=\"CaptionContent\"\u003e\n \u003cp\u003eClinicopathologic data of patients in relation to pCR and non-pCR status in the training and test sets\u003c/p\u003e\n \u003c/div\u003e\n \u003c/caption\u003e\n \u003ccolgroup cols=\"7\"\u003e\u003c/colgroup\u003e\n \u003cthead\u003e\n \u003ctr\u003e\n \u003cth align=\"left\"\u003e\u0026nbsp;\u003c/th\u003e\n \u003cth align=\"left\" colspan=\"3\"\u003e\n \u003cp\u003eTraining set (n\u0026thinsp;=\u0026thinsp;250) No. (%)\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\" colspan=\"3\"\u003e\n \u003cp\u003eTest set (n\u0026thinsp;=\u0026thinsp;108) No. (%)\u003c/p\u003e\n \u003c/th\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eCharacteristics\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003e(n\u0026thinsp;=\u0026thinsp;68)\u003c/p\u003e\n \u003cp\u003epCR\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003e(n\u0026thinsp;=\u0026thinsp;182)\u003c/p\u003e\n \u003cp\u003eNon-pCR\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003e\u003cem\u003eP\u003c/em\u003e value\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003e(n\u0026thinsp;=\u0026thinsp;29)\u003c/p\u003e\n \u003cp\u003epCR\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003e(n\u0026thinsp;=\u0026thinsp;79) Non-pCR\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003e\u003cem\u003eP\u003c/em\u003e value\u003c/p\u003e\n \u003c/th\u003e\n \u003c/tr\u003e\n \u003c/thead\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u003cstrong\u003eAge, years\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e47.2\u0026thinsp;\u0026plusmn;\u0026thinsp;10.6\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e50.2\u0026thinsp;\u0026plusmn;\u0026thinsp;11.9\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.346\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e48.5\u0026thinsp;\u0026plusmn;\u0026thinsp;12.3\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e51.1\u0026thinsp;\u0026plusmn;\u0026thinsp;13.0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.217\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u003cstrong\u003eTumor maximum size (mm)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e37.4\u0026thinsp;\u0026plusmn;\u0026thinsp;11.6\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e38.5\u0026thinsp;\u0026plusmn;\u0026thinsp;13.4\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.659\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e38.2\u0026thinsp;\u0026plusmn;\u0026thinsp;12.8\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e39.0\u0026thinsp;\u0026plusmn;\u0026thinsp;14.3\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.588\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u003cstrong\u003eClinical T stage\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.334\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.512\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eI\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e5 (7.4)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e7 (3.8)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e2 (6.9)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e3 (3.8)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eII\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e36 (52.9)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e88 (48.4)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e16 (55.2)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e37 (46.8)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eIII\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e27 (39.7)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e87 (47.8)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e11 (37.9)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e39 (49.4)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u003cstrong\u003eClinical N stage\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.273\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.392\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003ecN0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e16 (23.5)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e30 (16.5)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e6 (20.7)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e11 (13.9)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003ecN+\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e52 (76.5)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e152 (83.5)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e23 (79.3)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e68 (86.1)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u003cstrong\u003eHistologic type\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.497\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.685\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eInvasive ductal carcinoma\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e49 (72.1)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e123 (67.6)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e19 (65.5)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e55 (69.6)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eOthers\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e19 (27.9)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e59 (32.4)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e10 (34.5)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e24 (30.4)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u003cstrong\u003eER status\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e\u0026lt;\u0026thinsp;0.001*\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.002*\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003ePositive\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e26 (38.2)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e141 (77.5)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e10 (34.5)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e56 (70.9)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eNegative\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e42 (61.8)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e41 (22.5)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e19 (65.5)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e23 (29.1)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u003cstrong\u003ePR status\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e\u0026lt;\u0026thinsp;0.001*\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e\u0026lt;\u0026thinsp;0.001*\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003ePositive\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e22 (32.4)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e125 (68.7)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e8 (27.6)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e55 (69.6)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eNegative\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e46 (67.6)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e57 (31.3)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e21 (72.4)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e24 (30.4)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u003cstrong\u003eHER2 status\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e\u0026lt;\u0026thinsp;0.001*\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e\u0026lt;\u0026thinsp;0.001*\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003ePositive\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e40 (58.8)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e59 (32.4)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e20 (69.0)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e21 (26.6)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eNegative\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e28 (41.2)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e123 (67.6)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e9 (31.0)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e58 (73.4)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u003cstrong\u003eKi67 levels\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.523\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.369\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u0026le;\u0026thinsp;20%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e13 (19.1)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e43 (23.6)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e4 (13.8)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e17 (21.5)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u0026gt;\u0026thinsp;20%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e55 (80.9)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e139 (76.4)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e25 (86.2)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e62 (78.5)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u003cstrong\u003eMolecular subtypes\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e\u0026lt;\u0026thinsp;0.001*\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e\u0026lt;\u0026thinsp;0.001*\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eHR+/HER2-\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e11 (16.2)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e90 (49.5)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e4 (13.8)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e45 (56.9)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eHER2+\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e40 (58.8)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e59 (32.4)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e20 (69.0)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e21 (26.6)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eTriple-negative\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e17 (25.0)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e33 (18.1)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e5 (17.2)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e13 (16.5)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n \u003ctfoot\u003e\n \u003ctr\u003e\n \u003ctd colspan=\"7\"\u003eNote.\u0026mdash;There are 358 patients in the training and test sets. Mean data are \u0026plusmn;\u0026thinsp;standard deviation with the range. Data are the number of patients, with percentages in parentheses. ER\u0026thinsp;=\u0026thinsp;estrogen receptor, PR\u0026thinsp;=\u0026thinsp;progesterone receptor, HER2\u0026thinsp;=\u0026thinsp;human epidermal growth factor receptor 2, HR\u0026thinsp;=\u0026thinsp;hormone receptor, pCR\u0026thinsp;=\u0026thinsp;pathological complete response; *\u003cem\u003ep\u003c/em\u003e value\u0026thinsp;\u0026lt;\u0026thinsp;0.05.\u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tfoot\u003e\n \u003c/table\u003e\n \u003c/div\u003e\n \u003cp\u003eThe clinicopathologic model based on the factors associated with pCR in univariate analysis (i.e., ER, PR, and HER2 status, and molecular subtypes) yielded an AUC of 0.776 (95% CI: 0.717, 0.835), SEN of 78.6%, SPE of 82.1%, ACC of 76.9%, PPV of 68.3%, and NPV of 85.6% in the training set. In the test set, it obtained an AUC of 0.759 (95% CI: 0.657, 0.861), SEN of 74.6%, SPE of 79.3%, ACC of 74.8%, PPV of 66.5%, and NPV of 84.1%.\u003c/p\u003e\n \u003cp\u003eThe inter- and intraobserver reproducibility were substantial for the ROIs segmentation and radiomics features extraction, with ICCs\u0026thinsp;\u0026gt;\u0026thinsp;0.75 for all, and were robust for the further analysis.\u003c/p\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec10\" class=\"Section2\"\u003e\n \u003ch2\u003ePerformance of ML-based PURS and IURS models\u003c/h2\u003e\n \u003cp\u003eFor the PURS models, the RF classifier achieved the best predictive ability with an AUC of 0.882 (95% CI: 0.838, 0.923), than LR (AUC\u0026thinsp;=\u0026thinsp;0.866 [95% CI: 0.799, 0.921]), Adaboost (AUC\u0026thinsp;=\u0026thinsp;0.834 [95% CI: 0.760, 0.901]), SVM (AUC\u0026thinsp;=\u0026thinsp;0.738 [95% CI: 0.664, 0.819]), and LDA (AUC\u0026thinsp;=\u0026thinsp;0.687 [95% CI: 0.617, 0.761]) in the training set. In the test set, the AUCs were 0.889 (95% CI: 0.814, 0.947) for RF, 0.849 (95% CI: 0.734, 0.942) for LR, 0.823 (95% CI: 0.697, 0.933) for Adaboost, 0.746 (95% CI: 0.629, 0.848) for SVM, and 0.732 (95% CI: 0.625, 0.835) for LDA (Fig. \u003cspan class=\"InternalRef\"\u003e4\u003c/span\u003ea-e).\u003c/p\u003e\n \u003cp\u003eThe RF-based PURS predictive model also yielded favorable SEN of 80.8%, SPE of 87.4%, ACC of 85.6%, PPV of 70.5%, and NPV of 92.4% in the training set, and SEN of 75.9%, SPE of 89.7%, ACC of 85.9%, PPV of 73.3%, and NPV of 90.9% in the test set, respectively (Table\u0026nbsp;\u003cspan class=\"InternalRef\"\u003e2\u003c/span\u003e).\u003c/p\u003e\n \u003cdiv class=\"gridtable\"\u003e\u0026nbsp;\u003ctable id=\"Tab2\" border=\"1\"\u003e\n \u003ccaption language=\"En\"\u003e\n \u003cdiv class=\"CaptionNumber\"\u003eTable 2\u003c/div\u003e\n \u003cdiv class=\"CaptionContent\"\u003e\n \u003cp\u003eThe predictive performance of ML-based PURS, IURS models in the training and test sets\u003c/p\u003e\n \u003c/div\u003e\n \u003c/caption\u003e\n \u003ccolgroup cols=\"8\"\u003e\u003c/colgroup\u003e\n \u003cthead\u003e\n \u003ctr\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eTraining set\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eSEN (%)\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eSPE (%)\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eACC (%)\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003ePPV (%)\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eNPV (%)\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eAUC (95% CI)\u003c/p\u003e\n \u003c/th\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003ePURS models\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\u0026nbsp;\u003c/th\u003e\n \u003cth align=\"left\"\u003e\u0026nbsp;\u003c/th\u003e\n \u003cth align=\"left\"\u003e\u0026nbsp;\u003c/th\u003e\n \u003cth align=\"left\" colspan=\"2\"\u003e\u0026nbsp;\u003c/th\u003e\n \u003cth align=\"left\"\u003e\u0026nbsp;\u003c/th\u003e\n \u003cth align=\"left\"\u003e\u0026nbsp;\u003c/th\u003e\n \u003c/tr\u003e\n \u003c/thead\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eRF\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e80.8\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e87.4\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e85.6\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003e70.5\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e92.4\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.882 (0.838\u0026ndash;0.923)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eLR\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e77.9\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e93.4\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e89.2\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003e81.5\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e91.9\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.866 (0.799\u0026ndash;0.921)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eAdaboost\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e77.8\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e91.8\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e88.0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003e77.9\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e90.3\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.834 (0.760\u0026ndash;0.901)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eSVM\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e80.4\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e66.7\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e72.5\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003e49.9\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e93.8\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.738 (0.664\u0026ndash;0.819)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eLDA\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e69.1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e69.4\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e69.3\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003e45.6\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e85.8\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.687 (0.617\u0026ndash;0.761)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u003cstrong\u003eIURS models\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eRF\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e92.6\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e95.6\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e93.2\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003e88.7\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e97.2\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.948 (0.913\u0026ndash;0.978)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eAdaboost\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e89.7\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e77.6\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e81.8\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003e63.8\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e95.3\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.910 (0.873\u0026ndash;0.943)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eLR\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e73.5\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e81.4\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e79.3\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003e59.6\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e89.3\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.852 (0.800-0.897)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eLDA\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e67.7\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e81.9\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e80.7\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003e58.3\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e87.2\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.825 (0.757\u0026ndash;0.874)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eSVM\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e72.3\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e74.5\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e74.3\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003e59.2\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e85.6\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.817 (0.739\u0026ndash;0.868)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u003cstrong\u003eTest set\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u003cstrong\u003eSEN (%)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u003cstrong\u003eSPE (%)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u003cstrong\u003eACC (%)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003e\u003cstrong\u003ePPV (%)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u003cstrong\u003eNPV (%)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u003cstrong\u003eAUC (95% CI)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u003cstrong\u003ePURS models\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eRF\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e75.9\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e89.7\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e85.9\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003e73.3\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e90.9\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.889 (0.814\u0026ndash;0.947)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eLR\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e79.3\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e93.5\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e87.7\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003e82.1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e92.4\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.849 (0.734\u0026ndash;0.942)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eAdaboost\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e75.8\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e96.1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e90.6\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003e80.1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e91.4\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.823 (0.697\u0026ndash;0.933)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eSVM\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e68.9\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e85.9\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e81.3\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003e64.5\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e88.1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.746 (0.629\u0026ndash;0.848)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eLDA\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e72.4\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e73.0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e72.9\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003e50.0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e87.6\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.732 (0.625\u0026ndash;0.835)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u003cstrong\u003eIURS models\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eRF\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e96.5\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e93.5\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e94.3\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e88.8\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003e98.6\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.931 (0.865\u0026ndash;0.980)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eAdaboost\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e89.6\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e79.5\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e84.4\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e88.4\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003e95.3\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.920 (0.869\u0026ndash;0.967)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eLR\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e79.3\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e84.6\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e83.2\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e65.7\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003e91.6\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.875 (0.787\u0026ndash;0.935)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eSVM\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e72.4\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e84.4\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e81.3\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e64.9\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003e89.1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.825 (0.734\u0026ndash;0.904)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eLDA\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e75.0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e83.3\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e72.6\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e63.7\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003e88.2\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.798 (0.699\u0026ndash;0.887)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n \u003ctfoot\u003e\n \u003ctr\u003e\n \u003ctd colspan=\"8\"\u003eNote. \u0026mdash; SEN\u0026thinsp;=\u0026thinsp;sensitivity, SPE\u0026thinsp;=\u0026thinsp;specificity, ACC\u0026thinsp;=\u0026thinsp;accuracy, PPV\u0026thinsp;=\u0026thinsp;positive predictive value, NPV\u0026thinsp;=\u0026thinsp;negative predictive value, AUC\u0026thinsp;=\u0026thinsp;area under the receiver operating curve, CI\u0026thinsp;=\u0026thinsp;confidence interval, PURS\u0026thinsp;=\u0026thinsp;peritumoral ultrasound radiomics signature, IURS\u0026thinsp;=\u0026thinsp;intratumoral ultrasound radiomics signature, RF\u0026thinsp;=\u0026thinsp;random forest, AdaBoost\u0026thinsp;=\u0026thinsp;adaptive boosting, LR\u0026thinsp;=\u0026thinsp;logistic regression, SVM\u0026thinsp;=\u0026thinsp;support vector machine, LDA\u0026thinsp;=\u0026thinsp;linear discriminant analysis.\u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tfoot\u003e\n \u003c/table\u003e\n \u003c/div\u003e\n \u003cp\u003eTen, 7, 5, 3, and 2 optimal radiomics features were selected for the RF, LR, Adaboost, SVM, and LDA classifiers, respectively. The top three selected features in the RF classifier were wavelet-HHL_GLSZM_gray level non uniformity normalized (coefficient: 4.492), wavelet-LHH_GLRLM_run length non uniformity normalized (coefficient: 3.579), and wavelet-HLH_GLSZM_high gray level zone emphasis (coefficient: 1.700) (details are shown in Supplementary Appendix C-G).\u003c/p\u003e\n \u003cp\u003eFor the IURS models, the RF classifier obtained a maximum AUC of 0.948 (95% CI: 0.913, 0.978), compared with Adaboost (AUC\u0026thinsp;=\u0026thinsp;0.910 [95% CI: 0.873, 0.943), LR (AUC\u0026thinsp;=\u0026thinsp;0.852 [95% CI: 0.800, 0.897]), LDA (AUC\u0026thinsp;=\u0026thinsp;0.825 [95% CI: 0.757, 0.874]), and SVM (AUC\u0026thinsp;=\u0026thinsp;0.817 [95% CI: 0.739, 0.868]) in the training set. In the test set, the AUCs were 0.931(95% CI: 0.865, 0.980) for RF, 0.920 (95% CI: 0.869, 0.967) for Adaboost, 0.875 (95% CI: 0.787, 0.936) for LR, 0.825 (95% CI: 0.734, 0.904) for SVM, and 0.798 (95% CI: 0.699, 0.887) for LDA (Fig. \u003cspan class=\"InternalRef\"\u003e5\u003c/span\u003ea-e). The RF-based IURS predictive model also achieved a satisfactory SEN of 92.6%, SPE of 95.6%, ACC of 93.2%, PPV of 88.7%, and NPV of 97.2% in the training set, and SEN of 96.5%, SPE of 93.5%, ACC of 94.3%, PPV of 88.8%, and NPV of 98.6% in the test set, respectively (Table \u003cspan class=\"InternalRef\"\u003e2\u003c/span\u003e).\u003c/p\u003e\n \u003cp\u003eEleven, 8, 6, 4 and 2 optimal radiomics features were selected for the RF, Adaboost, LR, SVM, and LDA classifiers, respectively. The top three selected features in the RF classifier were wavelet-LHH_GLDM_large dependence low gray level emphasis (coefficient: 6.192), wavelet-LLL_GLCM_correlation (coefficient: 4.935), and wavelet-HHL_GLRLM_short run emphasis (coefficient: 2.302) (details are shown in Supplementary Appendix H-L).\u003c/p\u003e\n \u003cp\u003eDeLong test showed that both RF-based PURS and IURS models had higher ability than clinicopathologic model (0.882 vs. 0.776, Z\u0026thinsp;=\u0026thinsp;3.017, \u003cem\u003ep\u003c/em\u003e\u0026thinsp;=\u0026thinsp;0.001, and 0.948 vs. 0.776, Z\u0026thinsp;=\u0026thinsp;4.788, \u003cem\u003ep\u003c/em\u003e\u0026thinsp;\u0026lt;\u0026thinsp;0.001) in the training set, and (0.889 vs. 0.759, Z\u0026thinsp;=\u0026thinsp;3.646, \u003cem\u003ep\u003c/em\u003e\u0026thinsp;\u0026lt;\u0026thinsp;0.001, and 0.931 vs. 0.759, Z\u0026thinsp;=\u0026thinsp;4.059, \u003cem\u003ep\u003c/em\u003e\u0026thinsp;\u0026lt;\u0026thinsp;0.001) in the test set. Nevertheless, the RF-based PURS showed lower efficacy as compared with RF-based IURS (0.882 vs. 0.948, Z\u0026thinsp;=\u0026thinsp;2.970, \u003cem\u003ep\u003c/em\u003e\u0026thinsp;=\u0026thinsp;0.003, and 0.889 vs. 0.931, Z\u0026thinsp;=\u0026thinsp;2.247, \u003cem\u003ep\u003c/em\u003e\u0026thinsp;=\u0026thinsp;0.02) in both sets (Fig. \u003cspan class=\"InternalRef\"\u003e6\u003c/span\u003ea and b).\u003c/p\u003e\n\u003c/div\u003e"},{"header":"Discussion","content":"\u003cp\u003eReliable and noninvasive predictors of pCR may assist clinicians with precise NAC tactics for LABC patients. Previous researchers adopted breast cancer subtypes [\u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e22\u003c/span\u003e], US and mammographic images [\u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e23\u003c/span\u003e] to target NAC. More studies utilized MRI radiomics as a standardized image method to monitor treatment response [\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e, \u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e, \u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e]. A recent study reported that US radiomics compared favorably with MRI in the assessment of pCR [\u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e24\u003c/span\u003e]. However, it still remains unknown whether PURS may contribute to the response prediction. Also, the predictive value of various ML classifiers-based US radiomics has not been explored. The present study is the first attempt to assess the ability of various ML classifiers-based PURS, as compared with IURS and clinicopathologic factors to predict NAC effect. Our results showed that the RF classifier-based PURS exhibited a higher performance than clinicopathologic predictors, and a relatively low efficacy compare with IURS for the prediction of pCR (\u003cem\u003ep\u003c/em\u003e\u0026thinsp;\u0026lt;\u0026thinsp;0.05 for both) .\u003c/p\u003e \u003cp\u003ePeritumoral regions which contain a mixture of tumor cells and inflammatory elements have been reported to be associated with tumor aggressive and prognosis [\u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e25\u003c/span\u003e]. Braman et al. [\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e] applied intratumoral and peritumoral DCE-MRI radiomics for the pretreatment prediction of pCR to NAC. They defined 2.5- to 5mm radius surrounding the tumor as the peritumoral region, and yielded an AUC of 0.74 by using a combined set. Li et al. [\u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e26\u003c/span\u003e] reported that the ability to differentiate benign and malignant breast lesions within 3mm peripheral regions was similar to that of the entire internal regions on contrast-enhanced sonography (CEUS). Mao et al. [\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e] segmented five regions including intratumoral region, 5mm peritumoral region, 10mm peritumoral region, intratumoral\u0026thinsp;+\u0026thinsp;5mm peritumoral regions, and intratumoral\u0026thinsp;+\u0026thinsp;10mm peritumoral regions, and compared their performances on contrast-enhanced spectral mammography. Nevertheless, their study focused on the statistical difference between different areas of ROIs. Regions further away from the tumor boundary may involve more normal tissue, and embody less outcome-related information. In this study, we selected a 3mm-thick zone surrounding the tumor as the commonly used peritumoral region for PURS analysis, our results achieved a higher predictive efficacy than that of the combined models reported by Braman et al. [\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e]. However, as compared with IURS, the PURS models yielded a relatively lower predictive ability, which suggested that the immune response and the expansive lymphatic vessels with a relative preservation structure surrounding tumors may lead to resistance to treatment response.\u003c/p\u003e \u003cp\u003eRadiomics feature havs been regard as a potential predictor for clinical outcomes. In the present study, the dominant features selected in both PURS and IURS models were wavelet-related features. After the wavelet transform, GLSZM and GLRLM were the most selected features in PURS models, GLCM and GLDM were the mainly features in IURS models. GLRLM contains the length of the grayscale value parade, provides information on the spatial distribution of consecutive pixels in one or more directions. GLSZM is based on a principle similar to GLRLM, and can be calculated for the distances of different pixels or regions in the neighborhood. These features can indicate the comprehensive changed information of images in adjacent regions [\u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e27\u003c/span\u003e]. GLCM is a symmetric matrix representing the joint probability distribution of pixel pairs. GLDM includes the emphasis on large and small dependencies representing heterogeneity and homogeneity. Previous studies revealed that these signatures reflected the texture roughness and unevenness of an image, and indicated the intrinsic tumor heterogeneity [\u003cspan citationid=\"CR28\" class=\"CitationRef\"\u003e28\u003c/span\u003e]. Wavelet transform provides more valuable information about the tumor micro-environment through the recollection of texture features, with higher details and complexities than the original images [\u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e27\u003c/span\u003e, \u003cspan citationid=\"CR28\" class=\"CitationRef\"\u003e28\u003c/span\u003e]. Wavelet transform features has been proved to improve the diagnostic ability of malignant, the classification of breast cancer, and the prediction of prognosis or metastatic behavior [\u003cspan citationid=\"CR29\" class=\"CitationRef\"\u003e29\u003c/span\u003e, \u003cspan citationid=\"CR30\" class=\"CitationRef\"\u003e30\u003c/span\u003e]. Our results showed that the wavelet-related features of breast tumors were reliable predictors associated with pCR after NAC, which was compatible with previous studies [\u003cspan citationid=\"CR28\" class=\"CitationRef\"\u003e28\u003c/span\u003e, \u003cspan citationid=\"CR29\" class=\"CitationRef\"\u003e29\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eIn addition to ML algorithms, we used five popular classifiers (i.e., LDA, SVM, RF, LR, and AdaBoost) to construct PURS and IURS predictive models. The RF classifier achieved the most robust performance as compared with other classifiers in both PURS and IURS models. Tahmassebi et al. [\u003cspan citationid=\"CR31\" class=\"CitationRef\"\u003e31\u003c/span\u003e] applied eight ML classifiers with multiparametric MRI to predict pCR and survival outcomes. Their result showed that the XGBoost classifier achieved the most high accuracy for predicting residual cancer burden and disease-specific survival, and LR yielded a high AUC of 0.83 for predicting recurrence-free survival. However, their studies limited with only 38 patients, the small sample may be insufficient for clinical application. Sun et al. [\u003cspan citationid=\"CR32\" class=\"CitationRef\"\u003e32\u003c/span\u003e] indicated that RF classifier obtained the highest ability in the detection of TP53 mutations in TN and luminal type BCs. Masetic et al. [\u003cspan citationid=\"CR33\" class=\"CitationRef\"\u003e33\u003c/span\u003e] reported that the RF method yielded 100% classification accuracy in detecting congestive heart failure when compared with other classifiers. RF is a stochastic ensemble learning algorithm. With the combination of multiple decision trees at different subsets of data set, RF can effectively handle multitudinous high-dimensional data, and meanwhile avoid over-fitting [\u003cspan citationid=\"CR34\" class=\"CitationRef\"\u003e34\u003c/span\u003e]. Depending on the fast training speed, RF can quickly sort the importance of variables, and achieve the highest accuracy [\u003cspan additionalcitationids=\"CR33\" citationid=\"CR32\" class=\"CitationRef\"\u003e32\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR34\" class=\"CitationRef\"\u003e34\u003c/span\u003e]. Our results further corroborates the robustness of RF in the application of predicting treatment response.\u003c/p\u003e \u003cp\u003eIt is well-known that clinical and pathological factors are essential predictors for treatment effect [\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e, \u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e22\u003c/span\u003e]. Since the introduction of HER2-targeted drugs, such as trastuzumab and pertuzumab, HER2\u0026thinsp;+\u0026thinsp;has been regarded as a good prognostic factors of treatment response [\u003cspan citationid=\"CR35\" class=\"CitationRef\"\u003e35\u003c/span\u003e]. In contrast, the presence of hormone receptor (HR) positive has been reported as a poor prognostic factors of NAC [\u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e22\u003c/span\u003e]. In agreement with previous studies, our results showed that pCR to NAC occurred most commonly in HER2\u0026thinsp;+\u0026thinsp;patients, and the HR+/HER2- patients were the least sensitive to complete response [\u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e, \u003cspan citationid=\"CR35\" class=\"CitationRef\"\u003e35\u003c/span\u003e]. Previous studies also indicated that tumor size, tumor grade, and Ki67 levels were predictors associated with treatment effect [\u003cspan citationid=\"CR36\" class=\"CitationRef\"\u003e36\u003c/span\u003e]. However, in the present study, tumor size, clinical stage, histologic type, and Ki67 levels all had no significant difference between the pCR and non-pCR groups. The results suggested that the molecular subtypes of tumor may be more important than tumor size, stage, and proliferation in the determination of pCR status.\u003c/p\u003e \u003cp\u003eThe strengths of this study contributes to the fields of breast US radiomics in the following ways: First, it is the first attempt to explore the role of peritumoral environment based on US radiomics, in comparison with intratumoral region and clinicopathologic factors to predict pCR after NAC in LABC patients. Second, we applied five popular ML algorithms as a new approach to construct various predictive models, which may provide more convincing results. Third, compared to prior radiomics studies, which mainly focused on MRI or other image-based radiomics for monitoring NAC, US is a more convenient diagnostic tool in breast examination, thus making the US-based radiomics analysis with more wide application in clinical practice. Therefore, the present study represents a possible novel confluence of US radiomics to estimate peritumoral response, and is a further step toward \u0026ldquo;perfect\u0026rdquo; predictive tool to guide more individual treatment strategy.\u003c/p\u003e \u003cp\u003eCertainly, our study has some limitations: First, it was a retrospective study performed in a single institution. Second, we only included mass lesions with sufficient peritumoral tissue identified on US image, nonmass-like lesions and lesions with no clear peritumoral region on US images were excluded which may cause selected bias. In addition, we didn\u0026rsquo;t construct molecular subtypes models to assess the predictive performance because of the relative small number of subjects in each subgroup. Finally, our study did not include genomics data, but that was beyond the scope of our manuscript, further studies would be expected to address this issue in the future.\u003c/p\u003e \u003cp\u003eIn conclusion, although the RF classifier-based PURS yielded relatively lower efficacy than IURS model for the prediction of pCR, it also achieved a favorable predictive accuracy, with AUCs higher than clinicopathologic predictors. The consideration of PURS with ML algorithm as a novel and promising predictive tool may aid in clinical systemic therapy decision-making for LABC patients receiving NAC.\u003c/p\u003e"},{"header":"Abbreviations","content":"\u003cp\u003e\u003cstrong\u003eAdaBoost\u003c/strong\u003e\u003cstrong\u003e:\u003c/strong\u003e Adaptive boosting, \u003cstrong\u003eAUC:\u003c/strong\u003e Area under the curve, \u003cstrong\u003eBI-RADS:\u003c/strong\u003e Breast Imaging and Reporting Data System, \u003cstrong\u003eCI:\u0026nbsp;\u003c/strong\u003eConfidence interval, \u003cstrong\u003eER:\u003c/strong\u003e Estrogen receptor,\u0026nbsp;\u003cstrong\u003eGLCM:\u0026nbsp;\u003c/strong\u003eGray-level co-occurrence matrix, \u003cstrong\u003eGLDM:\u003c/strong\u003e Gray-level dependence matrix, \u003cstrong\u003eGLRLM:\u0026nbsp;\u003c/strong\u003eGray-level run-length matrix, \u003cstrong\u003eGLSZM:\u003c/strong\u003e Gray-level size-zone matrix,\u0026nbsp;\u003cstrong\u003eHER2:\u0026nbsp;\u003c/strong\u003eHuman epidermal growth factor receptor 2, \u003cstrong\u003eICC:\u003c/strong\u003e Intra-class correlation coefficient, \u003cstrong\u003eIURS:\u003c/strong\u003e Intratumoral ultrasound radiomics signatures, \u003cstrong\u003eLABC:\u003c/strong\u003e locally advanced breast cancer,\u0026nbsp;\u003cstrong\u003eLDA:\u0026nbsp;\u003c/strong\u003eLinear discriminant analysis,\u0026nbsp;\u003cstrong\u003eLR:\u003c/strong\u003e logistic regression, \u003cstrong\u003eNAC:\u003c/strong\u003e Neoadjuvant chemotherapy,\u0026nbsp;\u003cstrong\u003eNGTDM:\u003c/strong\u003e Neighboring gray tone difference matrix,\u003cstrong\u003e\u0026nbsp;PCR:\u0026nbsp;\u003c/strong\u003ePathological complete response,\u0026nbsp;\u003cstrong\u003ePURS:\u003c/strong\u003e Peritumoral ultrasound radiomics signatures,\u0026nbsp;\u003cstrong\u003eRF:\u0026nbsp;\u003c/strong\u003eRandom forest, \u003cstrong\u003eRFE:\u003c/strong\u003e Recursive feature elimination,\u0026nbsp;\u003cstrong\u003ePR:\u003c/strong\u003e Progesterone receptor, \u003cstrong\u003eROC:\u003c/strong\u003e Receiver operating characteristic curve, \u003cstrong\u003eSD:\u0026nbsp;\u003c/strong\u003eStandard deviation,\u0026nbsp;\u003cstrong\u003eSVM:\u0026nbsp;\u003c/strong\u003eSupport vector machine,\u0026nbsp;\u003cstrong\u003eUS:\u0026nbsp;\u003c/strong\u003eUltrasound\u003c/p\u003e"},{"header":"Declarations","content":"\u003ch2\u003eAuthor Contribution\u003c/h2\u003e\u003cp\u003eAuthors' contributions were as following: Jiejie Yao, Wei Zhou and Jianqiao Zhou had primary responsibility for the protocol development, patient enrolment, preliminary data analysis, and writing of the draft. Xiaohong Jia and Ying Zhu analyzed the data. Xiaosong Chen and Weiwei Zhan assisted with the data collection and verification. Jianqiao Zhou supervised data collection, and reviewed the manuscript for important intellectual content. Wei Zhou and Jianqiao Zhou supervised the design and execution of the study, contributed to the writing of the manuscript and had final approval of the manuscript submitted.All authors confirmed that they had full access to all the data in the study and accept responsibility to submit for publication.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\n\u003cli\u003eKorde LA, Somerfield MR, Carey LA, Crews JR, Denduluri N, Hwang ES, et al (2021) Neoadjuvant\u0026ensp;chemotherapy, endocrine\u0026ensp;therapy, and\u0026ensp;targeted therapy for breast cancer: ASCO Guideline. J Clin Oncol 39:1485-1505. doi: 10.1200/JCO.20.03399. \u003c/li\u003e\n\u003cli\u003eWang H, Mao X (2020) Evaluation of the efficacy of\u0026ensp;neoadjuvant\u0026ensp;chemotherapy\u0026ensp;for breast cancer. Drug Des Devel Ther.14:2423-2433. doi: 10.2147/DDDT.S253961. \u003c/li\u003e\n\u003cli\u003eChen S, Chen CM, Yu KD, Yang WT, Shao ZM (2012) A prognostic model to predict outcome of patients failing to achieve pathological complete response after anthracycline-containing neoadjuvant chemotherapy for breast cancer. J Surg Oncol. 105:577-85. doi: 10.1002/jso.22140. \u003c/li\u003e\n\u003cli\u003eM Braden A, V Stankowski R, M Engel J, A Onitilo A (2014) Breast\u0026ensp;cancer\u0026ensp; biomarkers: risk assessment, diagnosis, prognosis, prediction of treatment efficacy and toxicity, and recurrence. Curr Pharm Des. 20:4879-98. doi: 10.2174/1381612819666131125145517.\u003c/li\u003e\n\u003cli\u003eLiu Z, Li Z, Qu J, Zhang R, Zhou X, Li L, et al (2019) Radiomics of multiparametric MRI for pretreatment prediction of pathologic complete response to neoadjuvant chemotherapy in breast cancer: a multicenter study. Clin Cancer Res 25:3538-3547. doi: 10.1158/1078-0432.CCR-18-3190. \u003c/li\u003e\n\u003cli\u003eBian T, Wu Z, Lin Q, Wang H, Ge Y, Duan S, et al (2020) Radiomic signatures derived from multiparametric MRI for the\u0026ensp;pretreatment prediction\u0026ensp;of response to neoadjuvant chemotherapy in breast\u0026ensp;cancer. Br J Radiol. 93:20200287. doi: 10.1259/bjr.20200287.\u003c/li\u003e\n\u003cli\u003eHuang X, Mai J, Huang Y, He L, Chen X, Wu X, et al. Radiomic nomogram\u0026ensp;for\u0026ensp; pretreatment\u0026ensp;prediction\u0026ensp;of pathologic complete response to neoadjuvant therapy in\u0026ensp;breast\u0026ensp;cancer: predictive\u0026ensp;value of staging contrast-enhanced CT. Clin Breast Cancer 2021;21:e388-e401. doi: 10.1016/j.clbc.2020.12.004. \u003c/li\u003e\n\u003cli\u003eLee H, Lee DE, Park S, Kim TS, Jung SY, Lee S, et al (2019) Predicting response to neoadjuvant chemotherapy in patients with breast cancer: combined statistical modeling using clinicopathological factors and FDG PET/CT texture parameters. Clin Nucl Med. 44:21-29. doi: 10.1097/RLU.0000000000002348.\u003c/li\u003e\n\u003cli\u003eZhang MQ, Du Y, Zha HL, Liu XP, Cai MJ, Chen ZH, et al (2022) Construction and validation\u0026ensp;of a\u0026ensp;personalized nomogram of\u0026ensp;ultrasound for pretreatment prediction\u0026ensp;of breast cancer patients sensitive to neoadjuvant chemotherapy. Br J Radiol. 95:20220626. doi: 10.1259/bjr.20220626.\u003c/li\u003e\n\u003cli\u003eBraman NM, Etesami M, Prasanna P, Dubchuk C, Gilmore H, Tiwari P, et al (2017) Intratumoral\u0026ensp;and\u0026ensp;peritumoral\u0026ensp;radiomics\u0026ensp;for the pretreatment prediction of pathological complete response to\u0026ensp;neoadjuvant\u0026ensp;chemotherapy\u0026ensp;based on breast DCE-MRI. Breast Cancer Res.19:57. doi: 10.1186/s13058-017-0846-1.\u003c/li\u003e\n\u003cli\u003eMao N, Shi Y, Lian C, Wang Z, Zhang K, Xie H, et al (2022) Intratumoral and peritumoral radiomics for preoperative prediction of neoadjuvant chemotherapy effect in breast cancer based on contrast-enhanced spectral mammography. Eur Radiol 32:3207-3219. doi: 10.1007/s00330-021-08414-7. \u003c/li\u003e\n\u003cli\u003eHaug CJ, Drazen JM (2023) Artificial\u0026ensp;intelligence\u0026ensp;and\u0026ensp;machine\u0026ensp;learning\u0026ensp;in clinical medicine, 2023. N Engl J Med 388:1201-1208. doi.org/10.1056/NEJMra2302038.\u003c/li\u003e\n\u003cli\u003eYao J, Zhou W, Xu S, Jia X, Zhou J, Chen X, et al (2023) Machine learning-based breast tumor ultrasound radiomics for pre-operative prediction of axillary sentinel lymph node metastasis burden in early-stage invasive breast cancer. Ultrasound Med Biol. 10:S0301-5629(23)00335-6. doi: 10.1016/j.ultrasmedbio.2023.10.004.\u003c/li\u003e\n\u003cli\u003ePenault-Llorca F, Andr\u0026eacute; F, Sagan C, Lacroix-Triki M, Denoux Y, Verriele V, et al (2009) Ki67 expression and docetaxel efficacy in patients with estrogen receptor-positive breast cancer. J Clin Oncol 27:2809-2815. doi: 10.1200/JCO.2008.18.2808.\u003c/li\u003e\n\u003cli\u003eGradishar WJ, Moran MS, Abraham J, Aft R, Agnese D, Allison KH, et al (2022) Breast Cancer, Version 3.2022, NCCN Clinical Practice Guidelines in Oncology. J Natl Compr Canc Netw 20:691-722. doi: 10.6004/jnccn.2022.0030.\u003c/li\u003e\n\u003cli\u003eJones RL, Lakhani SR, Ring AE, Ashley S, Walsh G, Smith IE (2006) Pathological complete response and residual DCIS following neoadjuvant chemotherapy for breast carcinoma. Br J Cancer 94:358-362. doi: 10.1038/sj.bjc.6602950.\u003c/li\u003e\n\u003cli\u003eMercado CL (2014) BI-RADS update. Radiol Clin North Am 52:481-487. doi: 10.1016/j.rcl.2014.02.008.\u003c/li\u003e\n\u003cli\u003eFedorov A, Beichel R, Kalpathy-Cramer J, Finet J, Fillion-Robin JC, Pujol S, et al (2012) 3D Slicer as an image computing platform for the quantitative imaging network. Magn Reson Imaging 30:1323-1341. doi: 10.1016/j.mri.2012.05.001. \u003c/li\u003e\n\u003cli\u003eXu Z, Shen D, Kou Y, Nie T (2022) A\u0026ensp;synthetic \u0026ensp;minority\u0026ensp;oversampling \u0026ensp;technique based on gaussian mixture model filtering for imbalanced data classification. IEEE Trans Neural Netw Learn Syst. PP. doi: 10.1109/TNNLS.2022.3197156. \u003c/li\u003e\n\u003cli\u003eKernbach JM, Staartjes VE (2022) Foundations of machine learning-based clinical prediction modeling: Part II-generalization and overfitting. Acta Neurochir Suppl 134:15-21. doi:10.1007/978-3-030-85292-4_3.\u003c/li\u003e\n\u003cli\u003eSong Y, Zhang J, Zhang YD, Hou Y, Yan X, Wang Y, et al (2020) FeAture explorer (FAE): A tool for developing and comparing radiomics models. PLoS One 15:e0237587. doi:10.1371/journal.pone.0237587. \u003c/li\u003e\n\u003cli\u003evon Minckwitz G, Untch M, Blohmer JU, Costa SD, Eidtmann H, Fasching PA, et al (2012) Definition and impact of pathologic complete response on prognosis after neoadjuvant chemotherapy in various intrinsic breast cancer subtypes. J Clin Oncol. 30:1796-1804. doi: 10.1200/JCO.2011.38.8595. \u003c/li\u003e\n\u003cli\u003eSavaridas SL, Sim YT, Vinnicombe SJ, Purdie CA, Thompson AM, Evans A (2019) Are baseline ultrasound and mammographic features associated with rates of pathological completes response in patients receiving neoadjuvant chemotherapy for breast cancer? Cancer Imaging 19:67. doi: 10.1186/s40644-019-0251-3.\u003c/li\u003e\n\u003cli\u003eJiang M, Li CL, Luo XM, Chuan ZR, Lv WZ, Li X, et al (2012) Ultrasound-based deep learning radiomics in the assessment of pathological complete response to\u0026ensp;neoadjuvant\u0026ensp;chemotherapy\u0026ensp;in locally advanced\u0026ensp;breast cancer. Eur J Cancer 147:95-105. doi: 10.1016/j.ejca.2021.01.028. \u003c/li\u003e\n\u003cli\u003eCheon H, Kim HJ, Kim TH, Ryeom HK, Lee J, Kim GC, et al (2018) Peritumoral invasive breast cancer: prognostic value of peritumoral edema identified at preoperative MR imaging. Radiology 287:68-75. doi: 10.1148/radiol.2017171157. \u003c/li\u003e\n\u003cli\u003eLi J, Guo L, Yin L, Fang H, Ye W, Zhao B, et al (2018) Can different regions of interest influence the diagnosis of benign and malignant breast lesions using quantitative parameters of contrast-enhanced sonography? Eur J Radiol. 108:1-6. doi: 10.1016/j.ejrad.2018.09.005. \u003c/li\u003e\n\u003cli\u003eAbbasian Ardakani A, Bureau NJ, Ciaccio EJ, Acharya UR (2022) Interpretation of radiomics\u0026ensp;features - a pictorial review. Comput Methods Programs Biomed 215:106609. doi: 10.1016/j.cmpb.2021.106609. \u003c/li\u003e\n\u003cli\u003evan Griethuysen JJM, Fedorov A, Parmar C, Hosny A, Aucoin N, Narayan V, et al (2017) Computational radiomics system to decode the radiographic phenotype. Cancer Res 77:e104-e107. doi: 10.1158/0008-5472.\u003c/li\u003e\n\u003cli\u003eSudarshan VK, Mookiah MR, Acharya UR, Chandran V, Molinari F, Fujita H, et al (2016) Application of wavelet techniques for cancer diagnosis using ultrasound images: a review. Comput Biol Med. 69:97-111. doi: 10.1016/j.compbiomed.2015.12.006. \u003c/li\u003e\n\u003cli\u003eTang B, Chen Y, Wang Y, Nie J (2021) A\u0026ensp;wavelet-based learning model enhances molecular prognosis in pancreatic adenocarcinoma. Biomed Res Int. 7865856. doi: 10.1155/2021/7865856. \u003c/li\u003e\n\u003cli\u003eTahmassebi A, Wengert GJ, Helbich TH, Bago-Horvath Z, Alaei S, Bartsch R, et al (2019) Impact of machine learning with multiparametric magnetic resonance imaging of the breast for early prediction of response to neoadjuvant chemotherapy and survival outcomes in breast cancer patients. Invest Radiol. 54:110-117. doi: 10.1097/RLI.0000000000000518.\u003c/li\u003e\n\u003cli\u003eSun K, Zhu H, Chai WM, Yan FH (2023) TP53 mutation estimation based on MRI radiomics analysis for breast cancer. J Magn Reson Imaging 57:1095-1103. doi: 10.1002/jmri.28323. \u003c/li\u003e\n\u003cli\u003eMasetic Z, Subasi A (2016) Congestive heart failure detection using random forest classifier. Comput Methods Programs Biomed. 130:54-64. doi: 10.1016/j.cmpb.2016.03.020. \u003c/li\u003e\n\u003cli\u003eRigatti SJ (2017) Random\u0026ensp;Forest. J Insur Med. 47:31-39. doi: 10.17849/insm-47-01-31-39.1. \u003c/li\u003e\n\u003cli\u003eTakada M, Toi M (2020) Neoadjuvant\u0026ensp;treatment for HER2-positive\u0026ensp;breast\u0026ensp;cancer. Chin Clin Oncol. 9:32. doi: 10.21037/cco-20-123. \u003c/li\u003e\n\u003cli\u003eKeam B, Im SA, Park S, Nam BH, Han SW, Oh DY, et al (2011) Nomogram predicting clinical outcomes in breast cancer patients treated with neoadjuvant chemotherapy. J Cancer Res Clin Oncol. 137:1301-1308. doi: 10.1007/s00432-011-0991-3. \u003c/li\u003e\n\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":true,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true},"keywords":"Machine learning, Peritumoral and intratumoral ultrasound radiomics, Pathological complete response, Neoadjuvant chemotherapy","lastPublishedDoi":"10.21203/rs.3.rs-4440501/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-4440501/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003e\u003cstrong\u003ePurpose\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eNoninvasive, accurate and novel approaches to predict patients who will achieve pathological complete response (pCR) after neoadjuvant chemotherapy (NAC) could assist precise treatment strategies. The aim of this study was to explore machine learning (ML)-based peritumoral ultrasound radiomics signature (PURS), compared with intratumoral radiomics (IURS) and clinicopathologic factors, for early prediction of pCR.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eMethods\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eWe analyzed 358 locally advanced breast cancer patients (250 in the training set and 108 in the test set), who accepted NAC and post NAC surgery at our institution. The PURS and IURS of baseline breast tumors were extracted by using 3D-slicer and PyRadiomics software. Five ML classifiers including linear discriminant analysis (LDA), support vector machine (SVM), random forest (RF), logistic regression (LR), and adaptive boosting (AdaBoost) were applied to construct radiomics models for the prediction of pCR. The performance of PURS, IURS models and clinicopathologic predictors were assessed with respect to sensitivity, specificity, accuracy and the areas under the curve (AUCs).\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eResults\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eFor the PURS models, the RF classifier achieved better efficacy (AUC of 0.889) than LR (0.849), AdaBoost (0.823), SVM (0.746) and LDA (0.732) in the test set. For the IURS models, the RF classifier also obtained a maximum AUC of 0.931 than 0.920 (AdaBoost), 0.875 (LR), 0.825 (SVM), and 0.798 (LDA) in the test set. The RF-based PURS yielded higher predictive ability (AUC, 0.889; 95% CI: 0.814, 0.947) than clinicopathologic factors (AUC, 0.759; 95% CI: 0.657, 0.861; \u003cem\u003ep\u003c/em\u003e \u0026lt; 0.05), but lower efficacy compared with IURS (AUC, 0.931; 95%CI: 0.865, 0.980; \u003cem\u003ep\u003c/em\u003e \u0026lt; 0.05).\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eConclusion\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe peritumoral US radiomics, as a novel potential biomarker, may be a \u0026nbsp;promising clinical approach to guide precise therapy decisions.\u003c/p\u003e","manuscriptTitle":"Machine learning prediction of pathological complete response to neoadjuvant chemotherapy with peritumoral breast tumor ultrasound radiomics: compare with intratumoral radiomics and clinicopathologic predictors","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2024-06-07 20:31:30","doi":"10.21203/rs.3.rs-4440501/v1","editorialEvents":[{"type":"communityComments","content":1}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"29c74deb-d999-4bbc-bdf9-bb75b7e5fc23","owner":[],"postedDate":"June 7th, 2024","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[],"tags":[],"updatedAt":"2024-06-07T20:31:32+00:00","versionOfRecord":[],"versionCreatedAt":"2024-06-07 20:31:30","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-4440501","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-4440501","identity":"rs-4440501","version":["v1"]},"buildId":"zQwnuV7TCBrMSSSToR1PI","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}