Radiomics features for assessing tumor-infiltrating lymphocytes correlate with molecular traits of triple-negative breast cancer

We retrospectively enrolled 139 triple-negative breast cancer (TNBC) patients treated at the Fudan University Shanghai Cancer Center (FUSCC) from 1 August 2009 to 31 May 2015 with baseline dynamic contrast-enhanced magnetic resonance imaging (DCE-MRI) available who were suitable for radiomics analysis. In this TNBC radiomic cohort, transcriptomic data (n = 116) was also available. The framework of this study is presented as Fig. 1. The TNBC radiomics cohort (n = 139) was split into a training cohort (n = 98) and a validation cohort (n = 41) with a 7:3 ratio using a stratified randomization method to keep high and low TILs proportions similar in the two cohorts (Table 1). Quantification of stromal tumor-infiltrating lymphocytes (sTILs), fibrosis and necrosis were evaluated on pathological H&E staining area by two pathologists according to published guidelines [25, 26]. In this study, tumor-infiltrating lymphocytes (TILs) refer to sTILs unless otherwise specified. A percentage of sTILs ≥ 20% was defined as a high TILs level (Fig. 2A, B).

All the patients in this cohort underwent MRI with 1.5 T special breast magnetic resonance (Aurora Imaging Technology, Aurora Systems, Inc., Canada) and coils for breast. A series of cross-sectional images were obtained in prone position, including plain scan T2WI (TR 6680 ms, TE 68 ms, slice thickness 3 mm, slice spacing 1 mm), T1WI (TR 5 ms, TE 13 ms, slice thickness 3 mm, slice spacing 1 mm) and dynamic contrast-enhanced T1WI (TR 5 ms, TE 29 ms, slice thickness 1.1 mm, slice spacing 0 mm, FOV 360 × 360 mm). The contrast medium Gd-DTPA (0.2 mmol/kg, flow rate 2.0 ml/s) was injected 90 s after plain scan. Postcontrast images were obtained at 90, 180, 270, and 360 s after injection.

In this study, the tumor regions of interest (ROIs) were delineated semiautomatically on the peak enhanced phase of CE-MRI by 3D Slicer software (https://​www.​slicer.​org/​). ROIs were placed on all slices that contained the whole tumor or the largest lesion (in the case of multicentric or multifocal tumors). To ensure reproducibility, some of the ROIs were initially delineated by two radiologists at FUSCC (C.Y. and D.D.Z. with 9 and 4 years of experience in breast MRI, respectively). The inter- and intra-observer reproducibility of the ROIs and radiomic feature extraction were initially analyzed with the CE-MRI data of 60 randomly selected patients in a blinded fashion by two radiologists. Additionally, one radiologist (C.Y. with 9 years of experience in breast MRI) repeated the ROI drawing twice with an interval of at least 1 month and generated radiomic features following the same procedure. Intraclass correlation coefficients (ICCs) were utilized to evaluate the intra- and interobserver agreement in terms of feature extraction. Inter- and intraobserver reproducibility and radiomic feature extraction achieved substantial agreement with ICC > 0.75 both among the ROIs from the two radiologists and between the ROIs from the same radiologist [27]. An ICC greater than 0.6 was considered a marker of satisfactory inter- and intra-observer reproducibility. On the premise of good consistency, whole ROI segmentation was completed by the more experienced radiologist in each layer of the MRI scan.

All other phases were co-registered into the first postcontrast phase of DCE-MRI through non-linear registration using the symmetric normalization algorithm [28], which was performed in the ANTs toolbox, to eliminate the spatial mismatches caused by motion artifact. Nonparametric nonuniformity normalization (N3) algorithm was applied for bias field correction [29]. Moreover, Z-Score Normalization algorithm was used for data normalization.

We performed a feature extraction process on DCE-MRI images (four phases) based on the open source Pyradiomics package V3.0, implemented in Python 3.6 [30], including shape features, first-order features, textural features, wavelet domain features and time domain features. For spatial domain features, 14 shape-based features were common to all phases, which describe the difference in shape between different types of tumors. Eighteen first-order features and 75 textural features were calculated from the four phases individually. First-order features describe the distribution of voxel intensities, and textural features were obtained based on 5 textural matrices to describe the radiological pattern of the ROI, including the gray level cooccurrence matrix (GLCM), gray level dependence matrix (GLDM), gray level run length matrix (GLRLM), gray level size zone matrix (GLSZM), and neighboring gray tone difference matrix (NGTDM). Moreover, wavelet domain features were extracted for each first order feature and textural feature by applying wavelet filtering to the original images, yielding 8 decompositions per level. In addition, for time domain features, the extracted sequential features were mainly composed of the mean, variance, kurtosis and skewness of the time-varying curve constructed based on feature values in four phases, for each first order, textural and wavelet domain feature. The specific number of features and the corresponding calculation formulae are described in detail in the Additional file 1.

Elastic net regression and logistic regression were utilized to select the most predictive radiomics features from the extracted features and to train machine learning model, respectively. Specifically, in the training cohort, we selected the most predictive radiomics features characterizing TILs levels with elastic net regression [31]. Then, the logistic regression was performed with the selected features to develop a TILs prediction signature referred to as Rad-TILs. The probability of high TILs predicted by radiomics model (p) was generated by the following formula:

In this study, \(\beta 0\) = 0.9123828, \(\beta 1\) = −0.6522518, \(\beta 2\) = −0.9434133, \(\beta 3\) = −1.5792121, X1 = Enhanced-Phase-1-wavelet-LLL-Skewness, X2 = Skewness-wavelet-LLL-GLCM-IDMN, X3= Skewness-GLSZM-LGLZE. Rad-TILs score was defined as the probability of high TILs predicted by radiomics model (p), where higher Rad-TILs score indicated a higher predictive probability of high TILs based on the three representative radiomics features. The efficiency of the prediction model was assessed by the receiver operating characteristic (ROC), specificity, sensitivity and accuracy in the validation cohort.

The Rad-TILs score was calculated by the Rad-TILs model in the subcohort of patients whose RNA-seq data were available (n = 116). Patients were separated into high- and low-Rad-TILs groups by the median Rad-TILs score in the model training process. We conducted differentially expressed gene (DEG) selection (“limma” package in R) [32] and KEGG pathway analysis (“clusterProfiler” package in R) [33]. Furthermore, we conducted gene set enrichment analysis (GSEA) using the KEGG and Reactome databases (“clusterProfiler” package in R) [33] to compare enriched pathways between high- and low-Rad-TILs patients.

A previously published reference matrix of gene sets characterizing different immune cell populations suitable for breast cancer [34] was adopted in the present study. Single sample gene set enrichment analysis (ssGSEA) was used to calculate the immune cell abundance score in every patient (“GSVA” package in R) [35]. Then, the Wilcoxon test was utilized to compare the difference between the high- and low-Rad-TILs groups.

Cytokines, costimulators, coinhibitors and major histocompatibility complexes (MHCs) were compared between the two groups by transcriptomic analysis. Furthermore, two gene signatures characterizing T cell inflammation status [36] and T cell cytolytic activity [37] were adopted to infer the T cell status in the two groups of patients.

Student’s t test and Wilcoxon’s test were used to compare continuous variables. Prior to the comparisons, the normality of the distributions was tested with the Shapiro–Wilk test. Pearson’s chi-square test and Fisher’s exact test were employed for the comparison of unordered categorical variables. All the tests were two sided. P < 0.05 was regarded as indicating significance, and 0.05 < P < 0.1 was regarded as marginally significance unless otherwise stated. The false discovery rate (FDR) correction was used in multiple hypothesis testing to decrease false positive rates. All statistical analyses were performed with R software (version 4.0.3, http://​www.​R-project.​org).

We retrospectively curated 139 TNBC samples with preoperative DEC-MRI and post-operative H&E pathological slides to establish a TILs evaluation cohort. The intention of the study was split into two parts: generation of TILs prediction radiomics model and illustration of biological basis of the radiomics model (Fig. 1). A percentage of sTILs ≥ 20% was defined as high TILs level (Fig. 2A, B). With all the extracted radiomics features, we used elastic net regression to select the features that most closely correlated with tumor-infiltrating T lymphocytes (TILs) in the training cohort. The following three radiomics features were finally selected: Enhanced-Phase-1-wavelet-LLL-Skewness (spatial domain feature) that describes a first order imaging feature after applying wavelet filtering transformation of original first post-enhanced phase images, Skewness-wavelet-LLL-GLCM-IDMN (time domain feature) that depicts the variance pattern of a textural feature after wavelet filtering between each enhanced phase, and Skewness-GLSZM-LGLZE (time domain feature) that reflects the variance pattern of a textural feature between each enhanced phase (detailed description of radiomics features was presented in Additional file 1). The correlation between selected radiomics features and clinical characteristics was listed in Additional file 1: Table S1. In the training and validation cohorts, the three features presented a relatively lower value in tumors with high TILs (Fig. 2C). Then, the radiomic features were used as variables for logistic regression to build a prediction model. A predicted score reflecting the probability of high or low TILs (Radiomics TILs score, Rad-TILs score) for each patient was generated. We used the median of the Rad-TILs scores as a cutoff value to discriminate distinct Rad-TILs levels (Fig. 2D, E). The area under the receiver operating characteristic (ROC) curve (AUC) was 0.868 (95% CI 0.797–0.938) when predicting TILs in the training cohort, and the AUC was 0.790 (95% CI 0.638–0.943) in the validation cohort (Fig. 2F, G). In addition, the performance of the prediction model was tested for specificity (0.70), sensitivity (0.89) and accuracy (0.71) in the validation cohort.

Cohort of patient with available radiomics and RNA-seq data was then used to investigate the transcriptomic difference between two sets of patients with distinct Rad-TILs levels. First, differentially expressed genes (DEGs) related to immunity, such as CXCL11, CXCL13, and IDO1, were discovered between the two groups (Fig. 3A). KEGG analysis inferred that several pathways correlated with the immune response, such as antigen processing (p = 0.02) and presentation and PD-L1 expression and PD-1 checkpoint pathway in cancer (p = 0.04), were significantly upregulated in high Rad-TILs patients (Fig. 3B). From the most upregulated and downregulated pathways summarized by GSEA based on the KEGG and Reactome databases, we found that the upregulated pathways were mainly enriched in immune response and immune modulation. However, the downregulated genes were difficult to categorize (Fig. 3C, D). Moreover, natural killer cell-mediated cytotoxicity and T cell receptor signaling pathway were upregulated in high Rad-TILs patients (Fig. 3E). Thus, we verified the different immune responses at the transcriptome level between the high- and low-Rad-TILs groups predicted by radiomics.

We analyzed the correlation between Rad-TILs score levels and the clinicopathological characteristics of TNBC patients. High Rad-TILs score patients tended to have fewer pathologically positive lymph nodes (p = 0.073), but the difference was not statistically significant (Fig. 4A). Higher stromal TILs (sTILs) and immunohistochemistry (IHC) CD8 scores were detected in high Rad-TILs score patients (Fig. 4B, C). The intrinsic subtypes, mRNA subtypes, fibrosis and necrosis were equivalent between the two groups (Fig. 4D, E). Furthermore, tumor microenvironment (TME) cluster 3, which was proposed in our previous study to characterize the inflammatory immune status of TNBC [34], was significantly enriched in high Rad-TILs score patients (Fig. 4F).

We also compared the difference in the composition of immune cells in the TME inferred by RNA-seq between high- and low-Rad-TILs score patients. Memory B cells, M1 macrophages, activated NK cells, plasma cells, CD8 T cells, follicular helper T cells and regulatory T cells were significantly or marginally significantly increased in the high Rad-TILs score group (Fig. 5A). Based on two published immune cell signatures, we found that patients with high Rad-TILs scores exhibited higher cytolytic activity and T cell inflamed gene expression profiles (Fig. 5B, C).

A relatively inflammatory TME in high Rad-TILs tumors was also indicated by the comparison of key molecules on the cell surface and cell-cell interactions. Several molecules expressed on the cell surface, including costimulators, coinhibitors and major histocompatibility complex (MHC), were highly expressed in high Rad-TILs score patient tumors (Fig. 5D, E, J). In addition, the levels of secreted immune-related cytokines, such as interleukins (ILs), colony-stimulating factors (CSFs), interferons (IFNs) and chemokines, were significantly elevated in the high group of patients (Fig. 5F–I), while the levels of transforming growth factors (TGFs) and tumor necrosis factors (TNFs) were equivalent between the two groups. Consequently, we established the relationship between opaque radiomics features and meaningful molecular features. Apart from distinct TILs levels, high Rad-TILs TNBC samples exhibited a hot immune microenvironment.