A Radiomics Model Based on Synthetic MRI Acquisition for Predicting Neoadjuvant Systemic Treatment Response in Triple-Negative Breast Cancer

Abstract

Purpose

To determine if a radiomics model based on quantitative maps acquired with synthetic MRI (SyMRI) is useful for predicting neoadjuvant systemic therapy (NAST) response in triple-negative breast cancer (TNBC).

Materials and Methods

In this prospective study, 181 women diagnosed with stage I–III TNBC were scanned with a SyMRI sequence at baseline and at midtreatment (after four cycles of NAST), producing T1, T2, and proton density (PD) maps. Histopathologic analysis at surgery was used to determine pathologic complete response (pCR) or non-pCR status. From three-dimensional tumor contours drawn on the three maps, 310 histogram and textural features were extracted, resulting in 930 features per scan. Radiomic features were compared between pCR and non-pCR groups by using Wilcoxon rank sum test. To build a multivariable predictive model, logistic regression with elastic net regularization and cross-validation was performed for texture feature selection using 119 participants (median age, 52 years [range, 26–77 years]). An independent testing cohort of 62 participants (median age, 48 years [range, 23–74 years]) was used to evaluate and compare the models by area under the receiver operating characteristic curve (AUC).

Results

Univariable analysis identified 15 T1, 10 T2, and 12 PD radiomic features at midtreatment that predicted pCR with an AUC greater than 0.70 in both the training and testing cohorts. Multivariable radiomics models of maps acquired at midtreatment demonstrated superior performance over those acquired at baseline, achieving AUCs as high as 0.78 and 0.72 in the training and testing cohorts, respectively.

Conclusion

SyMRI-based radiomic features acquired at midtreatment are potentially useful for identifying early NAST responders in TNBC.

Keywords: MR Imaging, Breast, Outcomes Analysis

ClinicalTrials.gov registration no. NCT02276443

Supplemental material is available for this article.

© RSNA, 2023

See also the commentary by Houser and Rapelyea in this issue.

Summary

In participants with triple-negative breast cancer undergoing neoadjuvant systemic therapy, radiomic analysis applied to quantitative synthetic MRI data acquired after four cycles of therapy differentiated between those with and without pathologic complete response.

Key Points

• ■ A synthetic MRI acquisition applied to participants with triple-negative breast cancer before and after four cycles of neoadjuvant systemic therapy produced quantitative T1, T2, and proton density (PD) maps that were used for development of radiomic models for prediction of pathologic complete response (pCR).

• ■ Univariable analysis identified 15 T1, 10 T2, and 12 PD radiomic features at midtreatment that predicted pCR with an AUC greater than 0.70 in both the training and testing cohorts.

• ■ A multivariable radiomic model from T1 maps acquired at midtreatment predicted pCR with AUCs of 0.78 and 0.72 in the training and testing cohorts, respectively.

Introduction

Triple-negative breast cancer (TBNC) is a biologically aggressive subtype of breast cancer that lacks estrogen receptors, progesterone receptors, and human epidermal growth factor receptor 2. Patients with TNBC have an especially poor prognosis in comparison with patients with hormone-positive or human epidermal growth factor receptor 2–positive breast cancer. TNBC is associated with younger age, higher mitotic index, higher combined grade, and shorter life expectancy when compared with other molecular subtypes of breast cancer (1). Because of the lack of effective targeted therapies for TNBC, these patients are usually treated with neoadjuvant systemic therapy (NAST) (2), which is given before surgery to downstage the tumor and often consists of anthracycline and/or taxane-based chemotherapy (3,4), with or without immunotherapy (5). Less than half of patients with TNBC exhibit a pathologic complete response (pCR) to standard NAST (4), though the rate of pCR may increase to more than 60% with the addition of immunotherapy (6,7). Patients with TNBC with pCR have been reported to have a 5-year event-free survival rate of 90%, while up to 60% of patients with TNBC without pCR will develop local recurrence or distant metastasis within 5 years of initial diagnosis (8). Therefore, pCR has become a surrogate of excellent long-term outcome in patients with TNBC treated with NAST. However, pCR generally is not confirmed until surgical resection of the tumor after completion of NAST (3,9), which typically requires about 24 weeks. Early prediction of the treatment response can potentially triage patients with TNBC without pCR to alternative treatment regimens and spare them from the toxicity of ineffective therapy. Conversely, patients with TNBC predicted to achieve pCR with standard NAST may be considered for treatment de-escalation, reducing some unnecessary long-term exposure to treatment toxicity. Thus, there is an unmet need to develop noninvasive imaging biomarkers for accurate early prediction of NAST response in patients with TNBC, which may guide treatment strategies for improved outcomes.

Technical advances in MRI over the past decade have led to a method for the rapid and simultaneous quantification of tissue T1, T2, and proton density (PD), known as synthetic MRI (SyMRI) (10). SyMRI uses a saturation recovery-based fast spin-echo sequence that acquires data at multiple saturation delay times and echo times. The relaxation of spins across these delay and echo times causes signal changes that can be used to fit the quantitative spin parameters on a pixel-by-pixel basis. While most of the current commercial applications of SyMRI have been developed for brain imaging, the quantitative values acquired with SyMRI have been reported to determine prognostic factors in breast cancer (11–14). SyMRI does not require gadolinium-based contrast agents; however, the presence of gadolinium rim enhancement has been reported to distinguish between TNBC and non-TNBC lesions when assessed in combination with precontrast T1 (11) and T2 (15) values acquired with the SyMRI method. Histogram features such as 10th percentile, median, and 90th percentile values from histogram analysis of these quantitative values can distinguish breast cancer subtypes, including TNBC from luminal B subtypes (16). When performed with and without gadolinium-based contrast agents, SyMRI has also been reported to help predict NAST response in patients with TNBC (17).

Radiomic texture analysis incorporates hundreds to thousands of image-based features to characterize disease phenotypes (18–20). Previously, radiomic analysis has been used to distinguish between benign and malignant breast lesions by using breast MRI data (21,22), tumor classification (23–25), and prediction of treatment outcomes (26–30). These studies applied radiomic analysis on anatomic T1- or T2-weighted, diffusion-weighted, or dynamic contrast-enhanced images. While radiomic analysis is applicable to grayscale images produced by various medical imaging modalities, most MR images are affected by various scanner- and patient-dependent factors that can confound the relationship between pixel values and tissue properties. These factors include arbitrary pixel value scaling, transmit-and-receive field inhomogeneity, scanner hardware, and sequence and reconstruction parameter types and settings. On the other hand, quantitative imaging can account for these confounding factors to produce a measure that is more directly related to intrinsic tissue properties, which in turn may improve correlations between radiomic features and intrinsic tissue properties. To our knowledge, the application of radiomic analysis to the quantitative values produced by SyMRI in patients with TNBC undergoing NAST has not been investigated.

The aim of this study was to develop a model for predicting the response of TNBC to NAST by applying radiomic analysis to quantitative T1, T2, and PD values acquired with SyMRI before treatment and early during treatment.

Materials and Methods

Study Cohort

SyntheticMR provided their proprietary SyMRI software free of charge for research. The authors had control of the data and the information submitted for publication. A total of 195 participants with stage I–III TNBC were enrolled in the prospective clinical trial A Robust TNBC Evaluation FraMework to Improve Survival (ARTEMIS, ClinicalTrials.gov registration no. NCT02276443). In the current trial approved by our local institutional review board and performed in full accordance with the Health Insurance Portability and Accountability Act, participants were prospectively monitored for response to NAST at a single comprehensive cancer center. Written informed consent was obtained from all individuals before being enrolled in ARTEMIS. In this trial, all participants underwent MRI before treatment, at two time points during treatment, and after NAST. The ARTEMIS trial has led to multiple studies investigating the ability of several MRI techniques and analysis methods to help predict treatment response. For example, studies have assessed necrosis volume, percentage of contrast enhancement, apparent diffusion coefficient (31), amide proton transfer-weighted chemical exchange saturation transfer mapping (32), functional tumor volume at dynamic contrast-enhanced MRI (33), peritumoral apparent diffusion coefficient (34), and tumor-infiltrating lymphocyte levels with radiomics at dynamic contrast-enhanced MRI (35). None of these previous studies reported on the use of SyMRI or the radiomics model used in our study. From May 16, 2018, to August 17, 2020, 195 participants were imaged with SyMRI; 14 participants were excluded for lack of pretreatment MRI studies, presence of severe artifacts that obscured the majority of the primary tumor, lack of confirmed pCR status, or discontinuation of the ARTEMIS treatment protocol (Fig 1). The remaining 181 participants were assigned, consecutively by order of recruitment, first to the cross-validation training cohort (n = 119), followed by the independent testing cohort (n = 62). For multifocal or multicentric lesions, only one index lesion from each participant was analyzed in this study, resulting in a total of 181 lesions.

Figure 1:

A flowchart of the inclusion and exclusion criteria for the participants’ distribution in the training and validation cohorts. ARTEMIS = A Robust TNBC Evaluation FraMework to Improve Survival (ClinicalTrials.gov registration no. NCT02276443), C4 = after four cycles of neoadjuvant systemic therapy, pCR = pathologic complete response, SyMRI = synthetic MRI, TNBC = triple-negative breast cancer.

Pretreatment needle biopsy specimens were obtained at staging biopsy performed 2 or more weeks before baseline MRI for immunohistochemical assessment and histologic type. Estrogen receptor and progesterone receptor status were evaluated using immunohistochemical scoring, expressed as the percentage of cells with positive nuclear staining. TNBC was defined as a tumor that had less than 10% staining of invasive tumor cells at immunohistochemistry for estrogen receptor and progesterone receptor. Human epidermal growth factor receptor 2 was defined as negative per the American Society of Clinical Oncology–College of American Pathologists guidelines (36). Surgical specimens were assessed by dedicated breast pathologists. pCR was defined as the absence of residual invasive disease in the breast and axilla.

NAST consisted of dose-dense adriamycin (doxorubicin) and cyclophosphamide for four cycles over 8 (dose-dense) or 12 weeks, followed by paclitaxel every 2 weeks for four cycles or weekly for 12 doses. Thirty-one participants with suboptimal response or progression during adriamycin-cyclophosphamide chemotherapy were offered targeted therapy at clinical trials as the second phase of their NAST. After completing NAST, participants underwent surgical resection with assessment of residual disease by pathologic evaluation. Figure 2 shows a timeline of the treatment protocol, including the time points of the two MRI examinations.

Figure 2:

Timeline of treatment protocol and imaging time points. AC was given for 8 or 12 weeks, followed by paclitaxel over 12 weeks. Imaging was performed before and after the four cycles of AC, and pCR was confirmed with the surgical specimen. Radiomic features extracted from the SyMRI parameter maps acquired at MRI examinations were used to predict pCR. AC = adriamycin and cyclophosphamide, C4 = after four cycles of neoadjuvant systemic therapy, pCR = pathologic complete response, SyMRI = synthetic MRI.

MRI Data Acquisition

Breast MRI was performed with participants in the prone position in a 3-T whole-body scanner (MR750w; GE Healthcare) and with an eight-channel phased-array bilateral breast coil. Imaging was performed at baseline and after the initial four cycles of NAST (C4), using a protocol that included T1-weighted, T2-weighted, diffusion-weighted, and dynamic contrast-enhanced sequences. Axial bilateral SyMRI scans were acquired before administration of contrast media with the commercial MAGnetic resonance image Compilation (or, MAGiC; GE Healthcare) sequence, with the following sequence parameters: repetition time = 4000 msec, echo train length = 12, section thickness/spacing = 4 mm/1 mm, field of view = 340–360 mm, matrix = 320 × 256, acceleration factor = 2. The scan time for the SyMRI series was typically 6 minutes 8 seconds for 30 sections. T1, T2, and PD maps were generated from the source images by using SyMRI software (SyntheticMR).

Image Preprocessing, Phenotype-based Tumor Segmentation, and Volume Extraction

For phenotype-based tumor segmentation, three-dimensional contouring was performed on the breast tumor in the three parameter maps (T1, T2, and PD) at two different time points (baseline and C4). The tumor of interest was semiautomatically segmented, refined, edited, and modified section by section by two fellowship-trained breast radiologists with 15 and 7 years of experience (M.B. and R.M.M., respectively) by using in-house software (Image-I). One consensus region of interest was drawn per examination on the T2 parameter map or derived synthetic image, with further visual support from other sequences including a postcontrast three-dimensional T1-weighted sequence. A region drawn on the T2 map was automatically propagated to the other maps with Image-I. All segmentations were reviewed in consensus with senior breast imaging radiologists (G.M.R. and B.E.A., with 20 and 21 years of experience, respectively). Tumor biopsy marker clips and necrotic regions were excluded from the volume of interest to ensure proper phenotype border selection. The extracted volume of interest was calculated by multiplying the number of voxels in a region with the voxel volume.

Histogram and Radiomic Analysis

Radiomic features were extracted using an in-house source code based on MATLAB (MathWorks). Within the segmented volume of interest, a total of 310 measurements of 10 first-order histogram features and 300 second-order Haralick texture features (37–39) were extracted from each map at each time point (Fig 3). Histogram-based features obtained were minimum, maximum, mean, SD, skewness, kurtosis, and the first, fifth, 95th, and 99th percentile values. The texture features included 60 rotation-invariant gray-level co-occurrence matrix (GLCM) features calculated on images requantized to eight, 16, 32, 64, and 256 gray levels. For this analysis, mathematical relationships between co-occurring voxels separated by 1-voxel distance were evaluated in four in-plane directions. Twenty GLCM-based features were obtained per gray level: autocorrelation, contrast, correlation, cluster shade, cluster prominence, dissimilarity, energy, entropy, homogeneity, maximum probability, variance, sum average, sum variance, sum entropy, difference variance, difference entropy, information measure of correlation 1, information measure of correlation 2, inverse difference moment, and normalized inverse difference moment. Rotation invariant measures of the features were obtained by calculating the average, range, and angular variance of the GLCM values calculated in four in-plane directions. Thus, these three calculations were performed for five quantization levels and 20 GLCM features (40), producing a total of 300 Haralick radiomic features from a single image.

Figure 3:

Radiomic pipeline. (A) The segmentation process of T1, T2, and PD maps acquired using SyMRI. (B) Extracted volume of interest (VOI) after overlaying the segmented mask on the segmented maps. (C) Texture analysis process to extract different radiomic features. (D) Feature selective model to search for the best model to predict pCR and non-pCR status. pCR = pathologic complete response, PD = proton density.

Comparison of Demographic, Pathologic, and Volumetric Features

The outcome of treatment was pathologically evaluated in all 181 participants analyzed in the study. The participants were divided into two groups according to their response, pCR (n = 88) and non-pCR (n = 93). Of the 119 participants assigned to the training cohort, 62 had pCR and 57 did not; the 62 patients assigned to the testing cohort were evenly split between pCR and non-pCR.

Feature Selection, Model Building, and Identification of Radiomic Signature

Receiver operating characteristic curve analysis and area under the receiver operating characteristic curve (AUC) were used to evaluate the association and prediction accuracy for pCR outcome with the three maps generated from SyMRI scans (T1, T2, and PD) at the two time points (baseline and C4). At both time points, each participant had three parameter maps (PD, T1, and T2), and 300 GLCM and 10 first-order imaging features were computed for each parameter, producing a total of 930 measurements per participant at each time point. For univariable analysis, the AUC was calculated for each feature and parameter for the prediction of pCR status, where AUC = 0.5 was considered a random model. To build a multivariable prediction model, a logistic regression with elastic net regularization was performed for texture feature selection. The elastic net is a regularized regression method that linearly combines the penalties of the lasso and ridge methods (41), simultaneously realizing feature selection and model stability. For data assigned to the training set, both regularization and mixing parameters were optimized by using fivefold cross-validation based on AUC. Logistic regression models were then built using all features for combinations of quantitative parameters (T1, T2, PD) and time points (baseline, C4) to predict pCR using the selected radiomic features extracted from the volume of interest.

Statistical Analysis

Participant clinicopathologic characteristics were summarized using frequencies, percentages, means, SDs, medians, minimums, and maximums. Radiomic measurements were compared between pCR and non-pCR using the Wilcoxon rank sum test and Fisher exact test. The developed radiomic models were evaluated using data assigned to the independent testing set by calculating accuracy and the AUC with its associated P value. A P value less than .05 was considered statistically significant. The biostatistical analysis was done with R software (version 4.0.3; R Foundation for Statistical Computing) with packages caret (version 6.0–86) and pROC (version 1.16.2).

Results

Participant Characteristics and pCR Status

Participant demographic and clinical characteristics are summarized in Table 1. A total of 181 female participants were included in the study, with 119 (median age, 52 years [range, 26–77 years]) in the training cohort and 62 (median age, 48 years [range, 23–74 years]) in the testing cohort. There was no evidence of associations between demographic and clinical characteristics and pCR status (P > .05 for all). One hundred eighty participants were successfully scanned at baseline, with 118 from the training cohort and 62 from the testing cohort. Only 155 participants were successfully scanned at C4, with 105 from the training cohort and 50 from the testing cohort.

Table 1:

Demographic and Clinical Characteristics of Participants with Triple-Negative Breast Cancer in the Training and Validation Cohorts

Predictive Value of the Radiomic Model by Using Univariable Analysis

For the quantitative parameters PD, T1, and T2 extracted from the C4 time point, univariable analysis identified 12, 15, and 10 radiomic features, respectively, with an AUC greater than 0.70 in both training and testing cohorts for differentiating participants with versus those without pCR (Tables S1–S3). These top-performing radiomic features included entropy, variance, homogeneity, and energy. None of the parameters extracted at baseline nor any of the first-order histogram features demonstrated an AUC greater than 0.70 for both training and testing cohorts.

Predictive Value of the Radiomic Model by Using Multivariable Analysis

T1, T2, and PD radiomic-based multivariable models at baseline, C4, and combined baseline and C4 were evaluated to predict pCR status (Table 2). Multivariable models at C4 showed the best performance. Using elastic net regularization with fivefold cross-validation, the T2 model at C4 demonstrated the best performance on the training set. This radiomic signature predicted pCR status with AUCs of 0.79 and 0.67 (95% CI: 0.51, 0.83) in the training and testing cohorts, respectively. The T1 model at C4 demonstrated the best performance on the independent testing cohort. This radiomic signature was able to predict pCR status with AUCs of 0.78 and 0.72 (95% CI: 0.58, 0.87) in the training and testing cohorts, respectively. The multivariable combined PD, T1, and T2 model at C4 had the second-highest performance on the independent testing cohort, with AUCs of 0.76 and 0.70 (95% CI: 0.54, 0.86) in the training and testing cohorts, respectively. Models built on data acquired only at baseline or combined data acquired at both baseline and C4 did not achieve AUCs higher than 0.57 and 0.67, respectively, in the testing cohort.

Table 2:

Multivariable Analysis of Radiomic Features from the T1, T2, and Proton Density Maps at Different Time Points by Using Elastic Net Regularization with Fivefold Cross-Validation

Discussion

In this study, we evaluated the ability of radiomic features based on quantitative parameter maps acquired by SyMRI to noninvasively help predict pCR to NAST in participants with TNBC. We demonstrated that a radiomic signature based on these maps could be used to build a clinically useful predictive model that discriminates between pCR and non-pCR at the end of four cycles of NAST. Several individual GLCM features were found to differentiate the study participants with AUCs greater than 0.70 in both training and independent testing cohorts, including 12 features on the PD map, 15 features on the T1 map, and 10 features on the T2 map. Several top-performing individual features from our analysis have been previously reported to differentiate tumor characteristics and patient outcomes. These include features based on entropy and energy (sometimes referred to as uniformity), which are associated with heterogeneity and homogeneity within a region of an image (18,30), as well as angular variance calculations, which are associated with directional structures or edges (42). Two multivariable models (T1 and combined T1, T2, and PD) were able to predict pCR with an AUC of 0.70 or greater at C4 of NAST in TNBC. Our study shows the feasibility of a quantitative noncontrast MRI technique for prediction of the pCR status in patients with TNBC undergoing NAST. The parametric maps from SyMRI may also be used as part of a multiparametric imaging approach to pCR prediction. As factors affecting endogenous T1-, T2-, and PD-dependent signals are distinct from those of diffusion and perfusion imaging, the information provided by using SyMRI should be complementary to these other functional methods. Thus, a combination of SyMRI and diffusion imaging, with or without dynamic contrast-enhanced imaging, may potentially enable more accurate prediction of pCR from a single MRI examination.

While SyMRI provides the benefits of quantitative T1, T2, and PD imaging, its ability to simultaneously acquire this information in a single sequence provides further advantages over the acquisition of the three parameter maps separately with independent mapping techniques. The multiple SyMRI parameter maps are free of misregistration, and the shorter overall acquisition time of the SyMRI sequence decreases the risk of motion artifacts corrupting any single image set, which would otherwise corrupt a combined analysis of the maps. By using a single acquisition, SyMRI produces consistent voxel geometries for all parameter maps and generates the maps using a single unified fitting technique, while disparate mapping techniques based on different acquisition sequences and postprocessing methods may have individual requirements or limitations on voxel size, section spacings, or volume coverage. Finally, the rapid acquisition time makes multiple repetitions of the sequence more feasible, which may be necessary if a sequence needs to be repeated because of motion or artifact, or if a study's protocol were to perform SyMRI before and after administration of gadolinium-based contrast media. Although not included in our study, such data would allow additional derivation of parameters related to perfusion, and thus potentially allow a more complete characterization of the tissue. However, because SyMRI does not rely on any external contrast agents, promising diagnostic techniques based on quantitative T1, T2, and PD images may warrant further evaluation for patients with a low tolerance for gadolinium-based contrast agents. When performed without contrast media, the technique avoids potential issues and variability associated with various aspects of contrast agent administration, including human error, contrast kinetics, and differences between contrast agent formulations, with the added advantage of decreased cost. Gadolinium-based contrast agents have also been known to cause nephrotoxicity and allergic reactions (43); furthermore, while the long-term effects of gadolinium deposition in the nervous system are unknown (44), more caution is warranted for patients with TNBC undergoing NAST because they tend to be younger than patients with other forms of breast cancer (1) and may require serial imaging.

In breast cancer, SyMRI has been investigated as a potential synthetic imaging method (45) to produce quantitative T2 values similar to those acquired with multiecho spin-echo T2 mapping. Other studies have investigated the potential of differentiating prognostic factors or breast cancer subtypes by using SyMRI quantitative values (11–16), including one study differentiating TNBC from non-TNBC (15) and another study using histogram analysis on a variety of cancer subtypes (16). Matsuda et al (17) also investigated pre- and postcontrast SyMRI to assess NAST treatment response in a pilot study of 37 patients with TNBC. However, none of these have applied more advanced analyses, such as radiomic analysis or machine learning. Some of these studies have included diffusion (11,14), postcontrast T1 (15,17), or dynamic contrast-enhanced imaging (11,13) in their models to the benefit of relatively high AUC values (>0.80) for their differentiation tasks. Compared with previous studies, our study did not require any other imaging sequences for conducting analyses, was not reliant on gadolinium-based contrast media, included the largest sample size, and exclusively included patients with TNBC.

These same distinctions are true when comparing our work against studies that applied radiomics to MRI of breast cancer. To our knowledge, our study is also the first application of radiomic analysis to quantitative T1, T2, and PD data in patients with breast cancer. In MRI of breast cancer, the utility of radiomic analysis has been investigated for discrimination of malignant tumors (21,22), tumor classification and grading (24,25), prediction of NAST response (26–30), and prediction of recurrence, but all analyses were based on conventional T1-weighted and T2-weighted MRI and/or dynamic contrast-enhanced MRI data. Again, none of these studies used an exclusive cohort of individuals with TNBC. Of the studies aiming to predict NAST response (26–30), most had analyzed dynamic contrast-enhanced MRI studies. For standard-of-care MRI, the administration of contrast agents provides additional functional information about tumor perfusion, aiding study interpretation and treatment response prediction. Functional tumor volume measurements derived from dynamic contrast-enhanced MRI studies have been shown to be important predictors of NAST response and recurrence-free survival in patients with breast cancer (33,46–48). Studies applying radiomic analysis to contrast-enhanced MRI demonstrated the prediction of pCR with AUCs ranging from 0.74 (26) to 0.87 (28). Diffusion-weighted imaging with quantitative apparent diffusion coefficient analysis is another standard-of-care noncontrast MRI technique that has been evaluated for pCR prediction. Investigators have reported an AUC of 0.6 for change in apparent diffusion coefficient at midtreatment MRI for prediction of NAST response in all molecular subtypes of breast cancer and an AUC of 0.57 for TNBC in subgroup analysis (49). Radiomic models from the emerging noncontrast SyMRI technique were able to achieve AUCs greater than 0.7 at midtreatment on our cohort of participants with TNBC. While AUC in our multivariable analysis was as high as 0.79 using fivefold cross-validation in our training cohort, our study also verified this analysis on an independent data set, which was not included in any of the previously cited radiomic studies. The use of independent testing data serves to determine whether the parameters derived in the training or cross-validation set were overfitted. Most of the highest-performing classifiers did not use an independent testing set (eg, AUC of 0.91 [21]), likely because of a limited number of available individuals.

This study had several limitations. While necrosis could be identified on the parameter maps, the abundance of adipose tissue in the breast may obscure lesion borders, presenting the challenge of accurately contouring breast tumors. Because the SyMRI method does not suppress fat signal, some cases required our readers to contour on synthetic short τ inversion-recovery images generated by applying T1, T2, and PD values to a short τ inversion-recovery signal model, with further visual support from postcontrast three-dimensional T1 images. As the primary lesion had already been identified in these participants, the utility of this imaging and analysis technique was evaluated for therapy assessment rather than detection. The multiple image sets produced by using SyMRI may lend themselves to the development of an automated segmentation algorithm in future studies. Another limitation was that all imaging data for this study were acquired using a single scanner with a consistent protocol in individuals with TNBC. The uniformity of the study sample and acquisition methods reduced confounding factors in the data, but the resulting classifier may be specific to this acquisition technique. The performance of our model if applied to a different type of breast cancer, on a different scanner platform, or with different imaging parameters is unknown.

In summary, a radiomics-based model was applied to SyMRI data to predict pCR status in participants with TNBC undergoing NAST. Multivariable models composed of GLCM features on PD, T1, and T2 maps showed high performance differentiating between those with and without pCR in the independent testing cohort. This differentiation was achieved with imaging performed at midtreatment after four cycles of NAST, suggesting the potential to personalize treatment plans during therapy for improved outcomes. Advanced analysis techniques applied to quantitative parameter maps acquired by the noncontrast SyMRI method may be useful as an early response marker to NAST. Further work in a larger patient cohort may be needed to improve prediction accuracy of the radiomics-based model for clinical application.