Patients
The retrospective study was approved by the institutional review board of Fudan University Shanghai Cancer Center, and the need to obtain informed consent was waived. In this multicohort study, radiomics analysis was applied to three independent cohorts. A total of 328 women patients diagnosed with breast cancer histologically and TNBC immunohistochemically, and who received complete NAC with no prior treatments, underwent breast MRI before the start of NAC, and underwent surgery after NAC, were included in this study. The exclusion criteria included the following: (a) patients with a prior history of malignance (n = 8), (b) patients without pretreatment MRI or post-operative pathology (n = 23), (c) patients with poor qualities or motion artifacts on DCE-MRI (n = 4), (d) patients with marked BPE on DCE-MRI (n = 10), (e) and patients without obvious peritumoral vessel on DCE-MRI (n = 1) (Fig. 1). Finally, the dataset from our center between February 1, 2016 and May 31, 2019 was used as the primary cohort and consisted of 93 patients (mean age, 49 years; range 26–75 years). The dataset from our center between June 1, 2019 and February 26, 2021 was used as the internal validation cohort and consisted of 113 patients (mean age, 47 years; range 25–72 years). The other dataset from “Duke-Breast-Cancer-MRI” of The Cancer Imaging Archive (TCIA) [13] was used as the external validation cohort and consisted of 76 patients (mean age, 49 years; range 24–73 years).
In the primary and internal validation cohorts, ER, PR, HER2, Ki-67 index expression patterns, and axillary lymph node metastatic assessments were obtained from histopathologic reports of core biopsies performed before NAC administration. The immunohistochemical assessment of ER, PR, and HER2 was performed using the standard methods as previously reported [14]. In those tumors that were classified as 2+, HER2 genetic testing was confirmed by fluorescence in situ hybridization.
Neoadjuvant chemotherapy regimen and response assessment
In the primary and internal validation cohorts, the chemotherapy regimens included epirubicin/cyclophosphamide followed by docetaxel (EC followed by T), docetaxel/carboplatin (TCb), and EC. The median number of NAC cycles was six (range, 4–8). The mean interval between the end of NAC and surgery was 10 days (range, 3–27 days). There were no details of NAC regimens in TCIA cohort. pCR was determined by microscopic examination of the excised tumor and lymph nodes after the completion of NAC and defined as no invasive or noninvasive residual in breast or axillary nodes (ypT0 ypN0) [15].
MRI protocols
The detailed parameters of DCE-MRI acquisition of all cohorts can be found in Appendix E1 in the Supplementary Material. In the primary and internal validation cohorts, all breast MR examinations were performed within 14 days before the start of NAC. DCE-MRI was performed using a fat-suppressed T1-weighted 3D fast spoiled gradient-echo sequence before and five times continuously after a bolus injection of a gadolinium contrast agent (Magnevist, Bayer HealthCare Pharmaceuticals Inc.). The injections were performed with an automatic injector (OptiStar® Elite, Liebel-Flarsheim) at a dose of 0.1 mmol per kilogram of body weight and rate of 2 ml/sec, followed by a 20 mL saline flush. The subtraction and axial MIP images were generated automatically after acquisition.
In TCIA dataset, the contrast agents included Gadavist, Magnevist, and Multihanc with the volume of 10–20 ml. The subtraction and axial MIP images were manually calculated by the radiologist (xx, 5 years of experience).
Tumor Segmentation and Peritumoral Vessel Segmentation
All MR images were reviewed by two breast radiologists (TX, 5 years of experience; and QZ, 11 years of experience), who were blinded to the results of the treatment outcomes. For patients with multifocal or multicentric tumors, the tumors with the largest size and the ipsilateral vessel were segmented and analyzed on the basis of the axial MIP of the first postcontrast phase.
Tumor segmentation on the MIP image was conducted manually by the breast radiologist (TX, 5 years of experience). The region of interest (ROI) was delineated to include the entire tumor by using a free open-source software package (ITK-SNAP, version 3.8.0; http://itk- snap.org). If there was overlap between the index tumor and the peritumoral vessel in the axial MIP image, the intersection was removed using the eraser tool. The illustration for tumor segmentation can be found in Appendix E2 in the Supplementary Material.
The enhancement and segmentation of peritumoral vessel were performed by the eigenvalue analysis of the multiscale Hessian-based filter, which showed simultaneous noise and background suppression and vessel enhancement in MIP images [16]. The details of the multiscale Hessian-based filter method and peritumoral vessel segmentation by algorithm can be found in Appendix E3 and E4 in the Supplementary Material. The segmentation of peritumoral vessel were performed with the Python programming language (Scikit-image package, v. 3.6, Python Software Foundation, https://www.python.org/). Then, the peritumoral vasculature by algorithm segmentation was loaded to ITK-SNAP again, and a senior breast radiologist (xx, 11 years of experience) performed manual editing by painting missing voxels and erasing incorrect voxels to get the final peritumoral vasculature. The manual vessel editing procedure took approximately 4 minutes per case. The flowchart and illustration for the vessel segmentation procedures are shown in Fig. 2 and Fig. 3.
The final peritumoral vasculature, checked and edited by the breast radiologist (xx, 11 years of experience), represented the reference standard. To evaluate the performance of vessel detection by algorithm segmentation, the correct-detection rate, incorrect-detection rate, and missed-detection rate were computed (Appendix E5 in the Supplementary Material).
Radiomic feature extraction
After tumor and peritumoral vessel were segmented, the shape, statistical and textural features were extracted on MIP images using the PyRadiomics Python package [17]. For the tumor and peritumoral vessel detected on the MIP images, we extracted radiomics features, including 10 shape features, 19 first-order statistical features, and 70 texture features. Furthermore, we extracted 356 wavelet features (i.e., LL, LH, HL, HH) for each tumor. Wavelet features provide representative transformed domain information regarding intensity and textural features by decomposing the original image in low and high frequencies [18]. Finally, 455 features quantifying intratumoral characteristics and 99 features quantifying peritumoral vascular characteristics were obtained.
Radiomic feature selection and model development
All the radiomics features were scaled to a range of [0, 1] by using a minimum-maximum scaler. Then, the least absolute shrinkage and selection operator (LASSO) configured recursive feature elimination (RFE) method was applied to select features for the intratumoral model, peritumoral vascular model, individually. The k-nearest neighbor (k-NN) classifier was used to train and test the radiomics models for predicting pCR to NAC. The k-NN (k = 5) technique was trained based on the primary cohort, and then tested in the internal and external validation cohorts.
After building the tumor features-based prediction model and vessel features-based prediction model, an information fusion method was applied to fuse the prediction scores generated by the two models to improve the model performance [19]. The information fusion method included the minimum, maximum, and weighting average of the fusion.
Statistical analysis
Comparisons between the patient groups were employed with the Chi-square test or Fisher’s test for qualitative variables and the Student’s t-test or Mann-Whitney U test for quantitative variables. The areas under the receiver operating characteristic (ROC) curves (AUCs) were assessed and compared among the intratumoral model, peritumoral vascular model, and fusion model using the DeLong method [20]. Statistical analyses and radiomics analyses were performed with the Python programming language (v. 3.6, Python Software Foundation, https://www.python.org/). p\(<\)0.05 was considered statistically significant.