Introduction

Breast cancer, which is the most commonly diagnosed cancer globally, leads to more than 600,000 deaths from breast cancer annually [1]. It is a heterogeneous disease and is subcategorized into subtypes by hormone receptor (HR) and HER2 status; this includes luminal A, luminal B, HER2-positive, and triple-negative breast cancer (TNBC) [2]. Neoadjuvant chemotherapy (NAC) is often applied to patients with locally advanced breast cancer (LABC) to enhance breast-conserving surgery and downstaging of tumors [3]. Patients who achieve pathological complete response (pCR) after NAC could receive a survival benefit [4]. Hence, it is imperative to identify reliable biomarkers for NAC prediction. The I-SPY2 trial is a multicenter investigational drug trial for patients with LABC [5]. Furthermore, identifying patients who may respond to NAC with imaging and molecular information is also a critical task in I-SPY2. The primary efficacy endpoint in I-SPY2 is pCR.

Magnetic resonance imaging (MRI) is one of the most common technologies for breast cancer NAC response evaluation. In the I-SPY2 trial, dynamic contrast-enhanced MRI (DCE-MRI) was utilized at serial timepoints during NAC to adjust the randomization schema. Some previous I-SPY2 results have shown that MRI features, such as background parenchymal enhancement (BPE) and functional tumor volume (FTV), could be effective indicators to predict pCR [6,7,8,9]. In the past decades, with the advances in high-throughput technology, the molecular heterogeneity of breast cancer has been gradually recognized [10]. Analysis of corresponding molecular annotation of imaging phenotypes will not only provide a deeper understanding of these parameters but also boost their clinical application. Within the I-SPY2 program, gene expression arrays were used to select biomarkers that could predict pCR status [11]. The I-SPY2 trial and associated datasets provide an opportunity for the integrative analysis of MRI features and molecular phenotypes. These datasets provide molecular characterization, treatment arms, and MRI features information.

Integrative analysis medical imaging phenotypes and molecular profiles could identify biological basis of image features, which leads to improve their interpretability [12]. Furthermore, people want to identify imaging biomarkers that are associated with heterogeneity of tumors as substitutes for genetic testing. Hence, in this study, we aimed to integrate quantifying molecular biomarkers and MRI phenotypes to construct imaging-molecular relationships. Based on molecular phenotype information and quantified MRI features, we wish to achieve biological annotation of MRI features. Furthermore, clinical application values of biological feature-annotated MRI features also been explored.

Methods

Patient population

Patients at least 18 years old with a new diagnosis of stage II or III breast cancer and with a tumor size ≥ 2.5 cm were eligible and enrolled in the I-SPY2 trial [13, 14]. In this project, patients received control (12 weekly cycles of paclitaxel) or experimental treatment (12 weekly cycles of control treatment combined with experimental agents). Then, all patients received four cycles of anthracycline-cyclophosphamide (AC) treatment every 2–3 weeks. After NAC procedures, patients received surgical treatment. A total of 987 patients from 10 arms of I-SPY2 with gene expression profiles were included. In addition, 383 out of 987 also with corresponding MRI multi-features were used for analysis relationships between imaging phenotypes and molecular phenotypes.

Furthermore, a total of 1059 breast cancer patients in The Cancer Genome Atlas (TCGA) database with transcriptome data and clinical follow-up information were also downloaded from the UCSC XENA portal (GDC TCGA Breast Cancer, [10, 15]. This dataset was used to validate the prognostic impact of those genes related to MRI features. Furthermore, we also explored the prognostic impacts of infiltrating immune cells.

Deidentified data are available from a public database for scientific purposes. Approval was obtained from the institutional review board at each participating site. Written informed consent was also obtained from each patient at each participating site [5, 7, 11].

MRI feature analysis

MRI examinations were performed at four different timepoints in the course of NAC: T0 (pre-NAC), T1 (early NAC, after 3 cycles), T2 (mid-NAC, after 12 cycles), and T3 (post-NAC, before surgery). From The Cancer Imaging Archive (TCIA) database, 384 patients with four quantified DCE-MRI parameters, FTV, sphericity, BPE, and longest diameter (LD), were downloaded [7, 16,17,18]. Details of the MRI examination parameter calculation processes can be obtained from a previous study [7]. We obtained measured FTV, LD, BPE, and sphericity in I-SPY2 from the TCIA database.

A rectangular region of interest (ROI) encompassing the enhancing tumor region was determined by radiologist or experienced imaging coordinator. Tumor FTV masks were determined as the sum voxel volumes with percentage enhancement (PE) ≥ 70% and signal enhancement ratio (SER) ≥ 0. PE and SER were calculated using the following formula: PE = (S1–S0)/S0*100%. SER = (S1–S0)/(S2/S0)*100%. S0, S1, and S2 represent signal intensities at pre-contrast, early-, and late- postcontrast, respectively. Then, FTV and SPH were derived from the existing FTV masks [19].

LD values were obtained from the MRI reports at each site. LD is a required parameter in BI-RADS.

BPE represent the mean PE of fibroglandular tissue in the contralateral non-cancer breast. BPE was calculated in the contralateral breast. Target volume was set as the central 50% of axial sections in the breast. Fibroglandular tissue was identified from the target volume.

Sphericity is an index of the roundness of tumor region shapes relative to a sphere. Sphericity measurements range from zero to one. If sphericity is equal to 1, the shape of the tumor region is a perfect sphere.

Pretreatment molecular profiling

Core needle biopsies were acquired before treatment. Details were described in the previous study [11]. Briefly, samples were collected by using 16-gauge needle and stored at −80 ℃. Only samples contain at least 30% tumor cells that confirmed by pathologic examination were submitted to RNA extraction and gene expression quantification on the basis of Agilent platform. From NCBI’s Gene Expression Omnibus (GEO) database (https://www.ncbi.nlm.nih.gov/geo/, access ID: GSE194040), the gene expression matrix of 987 breast cancer patients in I-SPY2 was downloaded [11]. Gene expression levels were detected by GPL30493 (N = 333) and GPL20078 (N = 654).

Determination of the infiltration of immune cells

We obtained the marker gene sets for 24 immune cell types from a previous study [20]. We used the GSVA algorithm to derive the absolute enrichment scores for individual cancer samples. GSVA processes were performed by using the GSVA package in R software [21]. A total of 24 immune cells are involved in innate immunity: dendritic cells (DCs), immature DCs (iDCs), activated DCs (aDCs), eosinophils, mast cells, macrophages, natural killer cells (NKs), NK CD56dim cells, NK CD56bright cells, and neutrophils. And adaptive immunity: B cells, T helper 1 (Th1), Th2, T gamma delta (Tgd), CD8 + T, T central memory (Tcm), T effector memory (Tem), and T follicular helper (Tfh) cells.

Statistical analysis

Relationships between gene expression profiles and MRI features were identified based on Spearman correlation analysis. Positive (R > 0, P < 0.05) and negative (R < 0, P < 0.05) related genes for FTV, LD, BPE and sphericity were submitted to gene enrichment analysis, respectively. Gene enrichment analyses were performed by using the clusterProfiler package in R software to identify enriched Kyoto Encyclopedia of Genes and Genomes (KEGG) pathways [22, 23]. The clusterProfiler package calculates the enrichment test for KEGG pathways based on hypergeometric distribution. The area under the curve (AUC) of the receiver operating characteristic curves (ROCs) was calculated to validate the NAC prediction performance of those genes related to MRI features. Kaplan‒Meier (K-M) survival curves were generated to compare survival differences based on the median value of genes or MRI features. P < 0.05 was considered a significant difference unless otherwise specified.

Results

Patient dataset

In this study, two public datasets, I-SPY2 and TCGA, were included. We utilized I-SPY2 datasets for radiogenomics analysis and NAC response prediction estimation. MRI examinations were performed before treatment, during early treatment (after three cycles of therapy), at mid-treatment (12 cycles), and after all chemotherapy. Molecular profiling of pretreatment biopsy specimens was also acquired. TCGA dataset was used to validate clinical significance of some molecular factors. Our study schema is shown in Fig. 1.

Fig. 1
figure 1

I-SPY 2 study design and flowchart of this study. Patients were randomized to the control or one of the experimental drug arms. MRI examinations were performed at pretreatment, early-NAC, mid-NAC, and post-NAC. Available pretreatment MRI features and gene expression profiling were integrated for molecular annotation

Full size image

A total of 987 patients from 10 arms of I-SPY2 were included in our study. Of the 987 patients, 383 out of 987 patients were available with complete MRI multi-features and pretreatment molecular information was used to radiogenomics analysis (Fig. 2). A pCR was defined as the absence of residual invasive carcinoma in the breast and axillary lymph nodes, which confirmed by a surgical specimen after NAC. A total of 113 out of 383 patients achieved pCR, and 270 did not. The pCR rates varied in different treatment arms: control, 17.3% (14/81); trebananib, 20.7% (12/58); ganetespib, 36.6% (15/41); ganitumab, 17.8% (8/45); MK2206, 41.2% (14/34); neratinib, 20.0% (8/40); pembrolizumab, 48.1% (13/27); pertuzumab, 62.5% (10/16); T-DM1 + pertuzumab, 62.5% (9/17); veliparib/carboplatin, 41.7% (10/24). The pCR rates were 14.9% (24/161) for HR + /HER2 − , 31.7% (19/60) for HR + /HER2 + , 66.7% (20/30) for HR − /HER2 + , and 37.9% (50/132) for triple negatives (Fig. 3A).

Fig. 2
figure 2

Example MRI series of an I-SPY 2 patient who did not achieve pCR after NAC. The breast cancer patient was 65 years old. DCE-MRIs at four treatment timepoints T0 (A), T1 (B), T2 (C), T3 (D) were obtained from TCIA database (https://www.cancerimagingarchive.net/, project ID: I-SPY2 Trial)

Full size image
Fig. 3
figure 3

MRI multi-feature measurements and relationships between each other. A pCR rate in different treatment arms and different subtypes. B Measurements of four MRI features at different timepoints. C Relationships among MRI features

Full size image

The TCGA data portal, which provides 1069 breast cancer samples, was used to analyze the relationships between MRI feature-related genes and overall survival (OS). Detailed information about the patients’ clinicopathological characteristics is summarized in Table 1.

Table 1 Patient characteristics
Full size table

MRI multi-feature relationships

Four MRI features (FTV, LD, SPH, and BPE) were measured at four sequential timepoints (Fig. 3B). Percentage changes relevant to baseline (T0) value were calculated at each timepoint (ΔFTVT0_T1, ΔFTVT0_T2, and ΔFTVT0_T3; ΔLDT0_T1, ΔLDT0_T2, and ΔLDT0_T3; ΔSPHERICITYT0_T1, ΔSPHERICITYT0_T2, and ΔSPHERICITYT0_T3; ΔBPET0_T1, %ΔBPET0_T2, and ΔBPET0_T3) (Fig. 3B). Then, correlation relationships among these features were estimated. We found that FTV and LD measured at different timepoints were significantly correlated with each other. Sphericity measured at different timepoints is often negatively related to FTV and LD. Furthermore, high baseline BPE and sphericity measurements often related to a larger drop (Fig. 3C).

Molecular annotation for MRI multi-features

To identify potential molecular relationships between baseline MRI features and gene expression alterations, 383 patients with gene expression profiling and MRI features were integrated for analysis. Spearman correlation analysis was performed between each feature and the expression of 19,134 genes. Functional enrichment analysis showed that genes positively correlated with FTV were mainly enriched in some proliferation-related pathways. For example, the cell cycle is the most significant enriched pathway. Interestingly, genes negatively correlated with FTV were mainly enriched in immune-related pathways. LD positively correlated genes were significantly related to some signaling-related pathways. LD negatively correlated genes were also related to many immune-related pathways. Sphericity-related genes showed the opposite pattern of FTV and LD. Pathways positively related to sphericity mainly belonged to immune-related pathways, while pathways negatively related to sphericity were mainly enriched in cancer progression and invasive pathways. For example, genes positively related to sphericity most significantly enriched in antigen processing and presentation process. The number of BPE-related genes was lower than that of FTV-, LD- and sphericity-related genes. Hence, KEGG pathways related to BPE measurements were also less than that of FTV-, LD- and sphericity-related pathways. The top 10 most significant pathways for each parameter are summarized in Table 2.

Table 2 Top 10 significant Kyoto Encyclopedia of Genes and Genomes (KEGG) pathways enriched by MRI feature-related genes
Full size table

Considering that most MRI features were correlated with immune-related pathways, we also estimated the clinical significance of infiltrating immune cells and their relationships with MRI features. Increased enrichment scores of T cells, B cells, NK CD56bright cells, cytotoxic cells, pDCs, CD8 T cells, TFH cells, and NK cells were significantly correlated with superior OS (Fig. 4A). In the I-SPY2 molecular trial (N = 987), most immune cells were significantly more highly infiltrated in patients who achieved pCR. Eosinophils and Mast cells were higher in patients who without pCR (Fig. 4B). Spearman correlation analysis suggested that those MRI features with pCR prediction performance were most significantly related to immune cell infiltration status (Fig. 4C). For example, sphericity measured at T0 was positively correlated with Th1 cells, Tgd, T cells, NK CD56dim cells, cytotoxic cells, CD8 T cells, B cells, and aDC cells.

Fig. 4
figure 4

Clinical significance of tumor-infiltrating immune cells and MRI feature relationships. A Relationships between OS and 24 types of infiltrating immune cells in TCGA database; B differentially expressed immune cells between patients who achieved pCR and those who did not; C Spearman correlation relationships between MRI features and immune cell enrichment levels

Full size image

Clinical significance of MRI-feature-related genes

Spearman correlation analysis identified NCS1, CD83, ICAM3, and BFAR as the most significant genes related to FTV (positive), LD (negative), sphericity (positive), and BPE (positive), respectively. Therefore, we also estimated the prediction performance of these genes for pCR in the I-SPY2 molecular cohort (N = 987) and their prognostic impact on OS in the TCGA cohort (N = 1069). AUC results showed that NCS1 (AUC = 0.55, 95% CI:0.51–0.59, P = 0.009; Fig. 5A), CD83 (AUC = 0.63, 95% CI:0.59–0.66, P < 0.001; Fig. 5B), ICAM3 (AUC = 0.58, 95% CI:0.54–0.61, P < 0.001; Fig. 5C) all showed moderate performance for pCR prediction. However, BFAR did not show significant prediction performance (AUC = 0.53, 95% CI: 0.49–0.57, P = 0.158; Fig. 5D).

Fig. 5
figure 5

Clinical significance of the MRI multi-features most significantly related genes. ROC curves of NCS1 (A), CD83 (B), ICAM3 (C), and BFAR (D) for predicting NAC in I-SPY2 molecular cohort (n = 987); survival differences between high and low expression of NCS1 (E), CD83 (F), ICAM3 (G), and BFAR (H) based on median values in TCGA database

Full size image

In the TCGA database, we also explored the relationships between these genes and OS. Patients were divided into a high expression group or a low expression group based on the median expression value of each gene. Higher expression levels of NCS1 were correlated with poor OS in patients (Fig. 5E). Higher expression levels of CD83 was related to superior OS of patients and high expression levels of ICAM3 was trend significant related to OS (Fig. 5E, G). No survival differences were observed between the high BFAR and low BFAR groups (Fig. 5H).

Sphericity decrease and treatment response

Considering that %ΔSPHERICITYT0_T1 was also significantly negatively related to many immune signatures, its measurements may be effective for pCR prediction. To promote clinical application, sphericity decrease at the T1 timepoint was evaluated by a binary indicator of whether sphericity was decreased relative to T0. For example, sphericity was identified as a decreased status at T1 if %ΔSPHERICITYT0_T1 was less than 0 and non-decreased if %ΔSPHERICITYT0_T1 was more than or equal to 0. In the whole cohort, patients with sphericity-decreased status had a higher pCR rate when compared with those patients with sphericity-non-decreased status (P = 0.013). Similar results were also observed in the HR-positive HER2-negative cohort (P = 0.025). However, in the subcohort of HR-positive HER2-positive, HR-negative HER2-positive, and HR-negative HER2-negative patients, no significant pCR rate differences were observed (Table 3).

Table 3 Association between sphericity decrease and pathologic complete response
Full size table

Discussion

The ability of noninvasive approaches for breast cancer heterogeneity estimation and treatment response prediction has enormous clinical implications, as it could provide personalized clinical decision-making assistance. Integrative analysis of noninvasive MRI features and molecular profiling could provide a deeper understanding of image phenotypes and lead to better clinical application. Our results showed that MRI features were related to tumor molecular profiling alterations. Combined MRI features and gene expression profiling could be useful for effective clinical indicator identification.

Four MRI features (FTV, LD, sphericity, and BPE) were included in our study owing to their clinical significance. We integrated MRI features and gene expression profiling obtained at baseline. Among the four features we studied, FTV has investigational device exemption status and has been widely recognized in I-SPY 1 and 2 trials [24, 25]. Therefore, some previous studies have also attempted to explore the relationships between FTV and some pathological and/or molecular features. Xiao et al. studied the relationships between FTV and microvessel density (MVD). However, they did not identify FTV to be related to MVD [26]. Akin et al. demonstrated that FTV was strongly positively correlated with metabolic tumor volume by integrating MRI features and FDG-PET/CT analysis [27]. Here, we identified FTV-related genes and found that these genes mainly focused on cell proliferation-related pathways. Therefore, FTV may reflect the proliferation ability of breast cancer. LD is a parameter that was measured by a radiologist and reported in BI-RADS. We found that LD was also correlated with FTV. LD-positive related genes may be most enriched in signaling-related pathways. Sphericity ranges from 0 to 1, where a value equal to 1 indicates a perfect sphere. Therefore, sphericity measured in a solid round-shaped tumor is larger than in diffused tumors. We found that sphericity was positively correlated with immune-related pathways, while FTV and LD were negatively correlated with immune-related pathways. The number of BPE-related genes was lower than that of other features. BPE may not effectively capture the molecular characteristics of tumors. BPE was measured from contralateral non-cancer breast tissues. Therefore, BPE may reflect the hormone status of patients and NAC-induced physiologic changes in normal breast tissues.

The I-SPY 2 molecular project reported that responders have high levels of immune signatures [11]. Considering the great clinical significance of immune cells in breast cancer progression, we also explored the immune cell enrichment relationship with MRI features. Some studies have identified that immuno-phenotyping holds promise as an NAC prediction indicator and assists in novel therapeutic development [28,29,30]. We found that some features (sphericity, FTV, LD) related to infiltrating immune cells, whether positive or negative, were valuable in pCR prediction [31, 32]. In the I-SPY2 trial, most immune cells were more highly infiltrated in patients who achieved samples. These findings suggest that the immune microenvironment is the key molecular mechanism contributing to the NAC treatment response. Alterations in the immune microenvironment could be reflected by noninvasive MRI examination.

Sphericity-decreased status in T1 (early treatment) was related to many pCR-related molecular biomarkers which suggested his change status may be related to pCR. We analyzed whether sphericity was decreased at the individual patient level rather than how much it changes on average across patients. Specifically, associations between sphericity-decreased status and higher pCR rate may be explained by tumor morphological characteristics change that were sensitive to NAC. Solid tumors with higher sphericity measurements indicated more tend to spherical shapes, while scattered tumors have lower sphericity measurements. Previous study showed that sphericity at baseline or T1 was higher in patients with pCR when compared with patients who did not achieve pCR [32]. Our findings suggested that circumscribed lesions (patients with high sphericity values) had higher immune cells infiltrating status and more easily achieved to pCR. Furthermore, lesions that changed from solid to scattered status may indicated more response to NAC. Therefore, sphericity measurements and sphericity alteration status could act as pCR prediction indicators owing to it reflects the molecular characteristics of the tumor itself and its response to NAC.

Our radiogenomics analysis is also valuable in some clinical application fields. First, we found that some MRI features were significantly correlated to immune phenotypes of breast cancers. Immune phenotypes are effective biomarkers for immunotherapy in breast cancer [33, 34]. Hence, these MRI features may be potentially used for immunotherapy response prediction in the future. Second, some MRI or molecular biomarkers identified in our study could be used for NAC response prediction more interpretable.

However, this study has some limitations. First, only baseline gene expression profiling could be obtained. We did not analyze relationships between dynamic gene expression and MRI features. Dynamic gene expression could be more informative. Therefore, future studies focused on relationships between series MRI phenotypes and series gene expression profiles should be performed. Second, the sample size for analysis in this study was sufficient. However, patient distributions among different cancer subtypes or treatment arms were not even. Third, we only provided relationship annotations between images and molecular alterations, and causal relationships between them should also be explored. And preoperative study could provide more reliable information. Last, sphericity could not be calculated when FTV was close to zero in some patients who achieved pCR, which may influence stability of results.

In conclusion, our study provided relationships between MRI features and gene expression profiling. These findings could promote our further understanding of the biological characteristics of medical imaging phenotypes. Furthermore, integrative MRI features and gene expression profiling could be useful for the identification of reliable images or molecular indicators. Continued work to clarify the causal relationship between imaging phenotypes and molecular alterations is currently underway and promotes clinical application.