Machine learning radiomics of magnetic resonance imaging predicts recurrence-free survival after surgery and correlation of LncRNAs in patients with breast cancer: a multicenter cohort study

This study was conducted in accordance with the STROBE guideline checklist [11]. This study included three phases to train and validate the RDeepNet model for prediction of recurrence-free survival (RFS) and explore the association between radiomics and the treatment or epigenetic biological underpinning. In the RDeepNet model construction and validation phase (phase 1), the RDeepNet model was constructed with a combination of the intra- and peritumoral radiomic features using contrast-enhanced T1-weighted imaging (T1 + C) and T2-weighted imaging (T2WI) sequences, which aimed to pinpoint patients with a high or low risk of recurrence. The RDeepNet model was validated in an independent external validation cohort and a testing cohort. RNA-sequencing (RNA-seq) was performed to preliminarily explore the potential molecular mechanisms of radiomics. In phase 2, correlation and variance analyses were conducted to examine the changes of radiomics in patients before and after neoadjuvant chemotherapy with the response status. Based on the above findings, the association and quantitative relation of radiomics and epigenetic molecular characteristics were further analyzed with RNA-seq data in phase 3.

A total of 1,186 nonmetastatic invasive breast cancer patients were retrospectively recruited from four institutions in China, of which 73 patients did not pass the quality control (55 patients were not histologically confirmed to have stage I–III invasive breast cancer [12], and 18 patients lacked an MRI before surgery), and 1113 patients were finally enrolled. A total of 698 patients recruited from the national hospitals Sun Yat-sen Memorial Hospital of Sun Yat-sen University (Guangzhou, China) and Sun Yat-sen University Cancer center (Guangzhou, China) between March 23, 2011, and August 26, 2019, were assigned to a training cohort. Then, 171 patient cases collected from the Shunde Hospital of Southern Medical University (Foshan, China) and the Tungwah Hospital of Sun Yat-sen University (Dongguan, China) between March 09, 2012, and September 21, 2019, were used as the validation cohort. A total of 244 patients from the Sun Yat-sen Memorial Hospital of Sun Yat-sen University (Guangzhou, China) between April 19, 2013, and December 05, 2018, were assigned to the testing cohort. We retrospectively collected 92 formalin-fixed paraffin-embedded (FFPE) biopsy tissues from patients treated at the Sun Yat-sen Memorial Hospital of Sun Yat-sen University. All samples were reassessed by two pathologists and were found to contain more than 70% tumor cells. A total of 72 patients, who had both T1 + C and T2WI sequences from The Cancer Genome Atlas (TCGA) and The Cancer Imaging Archive (TCIA), were assigned to the TCGA cohort for assessing the efficacy of the deep learning prediction model.

The inclusion criteria were female patients aged at least 18 years with histological confirmation of stage I–III invasive breast cancer [12], underwent breast tumor and axillary MRI scans before surgery and axillary lymph node dissection, and who experienced perioperative therapy. Cases of patients with other previous or simultaneous tumors, incomplete pathological information, or unavailable standard MRI scans with or without contrast enhancement were excluded. The outcome was RFS, calculated from the date of surgery until the date of the most recent medical review or diagnosis of recurrence, or metastasis, and the association of radiomics with lncRNAs.

The four molecular subtypes of breast tumors were defined according to the St. Gallen Consensus Conference 2013 [13], with biomarkers measured by immunohistochemistry or in situ hybridization. Luminal A subtype patients were defined as estrogen receptor (ER)- and progesterone receptor (PR)-positive, HER2-negative, and Ki-67 level < 14%. Luminal B subtype patients were defined as ER-positive and over-expressed/amplified HER2, or ER-positive and HER2-negative, with Ki-67 level > 14%, or PR-negative/low. In contrast, ER- and PR-negative, HER2-positive subtype patients had over-expressed/amplified HER2, and triple-negative breast cancer (TNBC) subtype patients were HER2-negative.

Total RNA was extracted from FFPE samples using the QIAGEN FFPE RNeasy kit (QIAGEN GmbH, Hilden, Germany). RNA was analyzed using an Agilent RNA 6000 Nano Kit (Aglient Technologies, Santa Clara, CA, USA), and RNA integrity numbers were determined to evaluate RNA integration using an Agilent Bioanalyzer 2100 (Aglient Technologies, Santa Clara, CA, USA). An input of 500 ng of total RNA was amplified using an Ovation FFPE WTA System (NuGEN, San Carlos, CA, USA), and a NEBNext® Ultra™ II DNA Library Prep Kit (Illumina) was used for fragmentation and labeling. The quality and quantity of amplified libraries were evaluated using Qubit (Invitrogen, Carlsbad, CA, USA) and Agilent Bioanalyzer 2100 (Aglient Technologies, Santa Clara, CA, USA). All libraries were sequenced using a DNBSEQ-T7RS (MGI) with 100 bp paired-end reads. Base call files were converted to the fastq format using cal2Fastq. Raw data were normalized using the fastp (version 0.20.1) for data processing.

The acquisition protocol of the multiparametric MRI (including T1 + C, and T2WI) used across all institutions and the MR scanner parameters are described in Additional file 1: eAppendix 1 and Additional file 1: Table S1. All of the MRIs were normalized to obtain a standard normal distribution of image intensities using the N4ITK Bias Correction code. The 3D regions of interest (ROIs) in the breast intratumoral area and the peritumoral area (10-mm extension outward of the tumor parenchyma) were semi-automatically segmented using the 3D Slicer software (https://​www.​slicer.​org/​, version 4.10.2) [14]. The 3D regions of intra- and peritumoral (DICOM format) were transferred to the SlicerRadiomics code, a texture extraction platform based on the python package “PyRadiomics” [15]. For each patient, 3,452 quantitative radiomic features (863 features from each ROI in each sequence, including 12 diagnostic features, 107 original features, and 744 wavelet features) were extracted to analyze shape, size, intensity, morphology, and texture. Besides diagnostic features, the remaining radiomic features were categorized into seven groups: shape descriptors, first-order statistics, gray-level co-occurrence matrix (GLCM), gray-level size zone matrix (GLSZM), gray-level run-length matrix (GLRLM), gray-level dependence matrix (GLDM), and neighboring gray tone difference matrix (NGTDM). More details regarding the radiomic feature extraction are described in Additional file 1: eAppendix 2.

The Cox proportional hazards deep neural network, DeepSurv [8], was applied to construct the RDeepNet model for predicting individual recurrence risk. The network took 3,452 radiomic features as input for each patient. For the recurrence risk, the RDeepNet score was calculated with a single output node based on the negative log-partial likelihood function. The RFS predicted from the RDeepNet model was then assessed in the validation cohort and the testing cohort, respectively. More details about the network were described previously [8].

In total, 127 (52%) of the 244 patients from the testing cohort had radiomic features from before and after neoadjuvant chemotherapy, of which 72 (57%) patients were evaluated as responsive (complete response + partial response) to the therapeutic, with the standard of Response Evaluation Criteria in Solis Tumors (RECIST). The other 55 (43%) patients were defined as unresponsive (stable disease + progressive disease). The differential therapy-related radiomic features between responsive and unresponsive patients or before and after neoadjuvant chemotherapy were identified using the limma package, t test and paired samples t test, respectively. The heatmaps of the differentially expressed radiomic features were obtained with the R package pheatmap. The correlation matrix maps of the radiomic features extracted from intratumoral region were performed with the R package ggplots and RColorBrewer.

To explore the related biological mechanisms of radiomics, we performed RNA-seq for 92 patients from the training cohort. Additional file 1: Table S2 shows the clinicopathological characteristics of these patients. The compared files were downloaded from https://​www.​ensembl.​org/​index.​html and annotated with Perl software according to the ensemble ID of sequencing results. Next, the gene length was compared through the Gencode27 database on the basis of the counts data. Then, the counts data were converted into TPM data, and the lncRNAs were distinguished in accordance with the Ensembl database.

To explore the potential epigenetic biological underpinning of radiomics, 15 lncRNAs were selected using the Spearman’s rank correlation analysis and univariable Cox proportional hazards regression model in 92 patients with RNA-seq data. The limma package was utilized to identify the differential radiomic features between patients with high and low expression of the key lncRNA. The Gene Ontology (GO) and Kyoto Encyclopedia of Genes and Genomes (KEGG) analyses were performed using the clusterProfiler R package [16]. The pathways were also identified by running a gene set variation analysis (GSVA) with the R package gsva. The pathway enrichment analyses were considered statistically significant, with P values and false discovery rates of less than 0.05. Next, the deep learning prediction model of lncRNA expression was built with the intratumoral radiomic features based on the multilayer neural network (MLP) [17, 18]. A total of 92 patients with RNA-seq data were included for training the model, and 72 patients with both T1 + C and T2WI sequences from TCGA and TCIA were assigned to the TCGA cohort for assessing the efficacy of the model.

Fisher’s exact tests were performed to examine differences in the occurrence of categorical variables, while independent t tests were used to compare differences in continuous variables between the two groups. Survival was calculated using the Kaplan–Meier method and the log-rank test. Hazard ratios (HRs) and 95% confidence intervals (Cls) were calculated using a Cox regression analysis. Patients were categorized into high and low-risk groups with the optimal cutoff values defined by the R package ggsurvimier. The prognostic or predictive accuracy of the RDeepNet model and prediction model of lncRNA expression was assessed by using receiver operating characteristic curve (ROC) analysis. The performance of the RDeepNet model for RFS prediction and prediction model of lncRNA expression was evaluated by assessing sensitivity and specificity calculated by using the area under the ROC curve (AUC) method. For all analyses, two-sided P-values less than 0.05 were considered statistically significant. Statistical analyses were performed using R software (version 4.0.0).

This study included three phases to train and validate the RDeepNet model for prediction of RFS and explore the association between radiomics and the treatment or epigenetic biological underpinning, and we eventually achieved the prediction for expression of lncRNA with radiomic features based on deep learning. A total of 1113 patients from four academic institutions in China were eligible for this study (Additional file 1: Table S3). Additional file 1: Table S4 shows the clinicopathological characteristics of patients in the training cohort (n = 698), the validation cohort (n = 171), and the testing cohort (n = 244). Endocrine therapy was administered to 446 (64%) of 698 patients in the training cohort, 103 (60%) of 171 patients in the validation cohort, and 135 (55%) of 244 patients in the testing cohort. HER2-targeted therapy was administered to 210 (30%) of 698 patients in the training cohort, 50 (29%) of 171 patients in the validation cohort, and 93 (38%) of 244 patients in the testing cohort. From the testing cohort, 244 patients underwent neoadjuvant chemotherapy. The median follow-up time was 44.7 months (IQR 34.0–57.3) for the training cohort, 40.4 months (IQR 29.3–62.3) for the validation cohort, and 39.9 months (IQR 36.1–50.9) for the testing cohort. The 3-year RFS rate was 93.6% (95% CI 91.7–95.5%) for the training cohort, 96.7% (95% CI 93.8–99.6%) for the validation cohort, and 93.3% (95% CI 90.0–96.6%) for the testing cohort. Detailed information regarding the patient recruitment and study design is described in Fig. 1.

The RDeepNet model, combining both intratumoral and peritumoral radiomic features, was developed. The RDeepNet model categorized patients into high- and low-risk groups with an optimal cutoff value (1.10). The RDeepNet model assigned 70 (10%) of 698 patients to the high-risk group, and there were significant differences in RFS between the high- and low-risk groups (HR 0.03, 95% CI 0.02–0.06, P < 0.001). In the validation cohort, 26 (15%) of the 171 patients were assigned to the high-risk group, which had shorter RFS (HR 0.05, 95% CI 0.01–0.23, P < 0.001). In the testing cohort, 59 (24%) of the 244 patients with high risk had shorter RFS (HR 0.05, 95% CI 0.02–0.19, P < 0.001) (Fig. 2a–c). Moreover, the RDeepNet model showed AUCs for the 1-, 2-, and 3-year RFS of 0.98, 0.94, and 0.92, respectively, in the training cohort; 0.91, 0.90, and 0.91, respectively, in the validation cohort; and 0.92, 0.93, and 0.94, respectively, in the testing cohort (Fig. 2d–f).

In addition, the RDeepNet model was employed to classify a high and low risk of recurrence in patients by considering the molecular subtypes of cancer. Encouragingly, the RDeepNet model could discriminate high- from low-risk patients in the subgroups of luminal A (P < 0.001), luminal B (HR 0.06, 95% CI 0.03–0.10, P < 0.001), HER2-positive (HR 0.05, 95% CI 0.01–0.22, P < 0.001), and TNBC (P < 0.001) patients (Additional file 1: Fig. S1). Moreover, the RDeepNet model could recognize high- and low-risk patients among patients treated with endocrine therapy (HR 0.03, 95% CI 0.02–0.07, P < 0.001) and patients treated with HER2-targeted therapy (HR 0.07, 95% CI 0.03–0.14, P < 0.001) (Fig. 3a, b). In parallel, the efficacy of the RDeepNet model showed AUCs of 0.95, 0.93, and 0.90 for 1-, 2-, 3-year RFS prediction among patients treated with endocrine therapy. These AUCs were 0.96, 0.90, and 0.90, respectively, among patients treated with HER2-targeted therapy (Fig. 3c, d).

According to the RDeepNet model, the radiomic features were expressed differentially between the high- and low-risk groups among patients in the training cohort (Additional file 1: Fig. S2). To determine the potential mechanisms of radiomics, RNA-seq for 92 patients from the training cohort was performed. We identified 148 differentially expressed genes between the high- and low-risk groups (Additional file 1: Fig. S3a). The pathway enrichment analyses showed that these genes were highly enriched in PI3K-Akt signaling pathway, MAPK signaling pathway, and cell-migration-related genomic biological processes (Additional file 1: Fig. S3b, c). These genes were also involved in various pathways as well as physiological and pathological processes, which were associated with tumor, immunity and metabolism, such as JAK STAT signaling pathway, cytokine interaction, and the energy metabolism (Additional file 1: Fig. S3d). We further evaluated the association between the RDeepNet score and immune cells (Additional file 1: Fig. S4a). Correlation analysis indicated that the RDeepNet score was significantly related to activated dendritic cells, CD56bright natural killer cells, central memory CD4 T cells, effector memory CD4 T cells, effector memory CD8 T cells, myeloid-derived suppressor cell, T follicular helper cells, Type 1 T helper cells, Type 17 T helper cells, and Type 2 T helper cells (Additional file 1: Fig. S4b). Furthermore, variation analysis of immune cells showed that patients from the high-risk group had lower expression of CD56dim natural killer cells and central memory CD8 T cells but higher expression of effector memory CD8 T cells compared with low-risk patients (Additional file 1: Fig. S4c).

After the RDeepNet model construction and validation phase, in phase 2, we aimed to determine whether radiomic features can predict changes after neoadjuvant chemotherapy. Radiomic variation and correlation analyses were performed on 127 patients (including 72 responsive patients and 55 unresponsive patients) who had intratumoral radiomic features, both before and after neoadjuvant chemotherapy. A total of 1726 intratumoral radiomic features were analyzed with consideration of patients’ response to the therapy. Before neoadjuvant chemotherapy, 456 radiomic features were found to be differentially expressed between responsive and unresponsive patients (Fig. 4a). It was observed that there were 352 variant radiomic features between the above two groups of patients after neoadjuvant chemotherapy (Fig. 4b). In addition, 306 and 793 radiomic features were found to be statistically different after the neoadjuvant chemotherapy in the responsive and unresponsive patients, respectively (Fig. 4c, d). The correlation between these features changed obviously after neoadjuvant chemotherapy. Patients after neoadjuvant chemotherapy had higher correlations among some radiomic features than patients before neoadjuvant chemotherapy (Fig. 4e–h). We further took the overlaps of the above differential radiomic features, and 35 radiomic features (therapy-related features) were considered to be the key features that were primarily correlated with therapy (Additional file 1: Fig. S5). It is worth noting that 27 of the 35 radiomic features were found to be significantly different between the high- and low-risk groups. Most of the key differential radiomic features were found to belong to the classification of the GLCM or GLRLM. More details about the classification of these features are shown in Additional file 1: Table S5.

Based on the above findings, the association of radiomics and epigenetic molecular characteristics was explored in phase 3 based on the results of RNA-seq. A total of 12,312 lncRNAs were produced from the transcriptome sequencing data for each patient. To look for the crucial lncRNAs, correlation analysis between lncRNAs and the RDeepNet score was performed, which allowed the identification of 47 lncRNAs at a significance threshold of P < 0.05. Of these, 15 were associated significantly with RFS (All P < 0.05) (Additional file 1: Figs. S6–8). The full list of the 15 lncRNAs is detailed in Additional file 1: Table S6. The association among the risk stratification by the RDeepNet model, pathological tumor-node-metastasis (pTNM) stage, molecular subtypes, and the 15 lncRNAs is visualized in Fig. 5a. Five (KRT7-AS, DLGAP1-AS2, AP000253.1, AC073130.2, LINC00910) of 15 lncRNAs were identified to be highly correlated with some of the above 35 therapy-related radiomic features (Fig. 5b).

The lncRNA KRT7-AS in particular was observed to be associated with RFS (HR 0.12, 95% CI 0.030–0.52, P < 0.001) (Fig. 6a), and was correlated linearly with most of the therapy-related radiomic features. Similarly, 22 of 35 therapy-related radiomic features were found to be differentially expressed in patients with differential expression of lncRNA KRT7-AS (Fig. 6b, c). To explore the potential biological underpinning, pathway enrichment analysis was conducted to evaluate the enrichment of the lncRNA KRT7-AS-related and survival-based genes. Figure 6d shows the lncRNA KRT7-AS mainly associated with various tumor- or metastasis-associated pathways and processes, such as the Akt phosphorylates, nucleotide excision repair, and ERBB2 regulates cell motility. The lncRNA KRT7-AS was also found to be correlated with effector memory CD8 T cells, immature dendritic cells, myeloid-derived suppressor cells, monocytes, neutrophils, type 1 T helper cells, and type 17 T helper cells. The results show that genes based on different expression of KRT7-AS were involved in the process of lncRNA-mediated mechanisms of therapeutic resistance, the Hippo-YAP signaling pathway, and TP53 network, which was also associated with tumors and survival (Fig. 6e). These results are consistent with previous research findings that the lncRNA KRT7-AS could promote tumor progression [19‐21].

The above findings show that the lncRNA KRT7-AS was obviously associated with radiomics and mediated in the progression of breast cancer. They remind us that it was feasible to predict the expression of KRT7-AS using radiomics. A deep learning prediction model was constructed based on the MLP among the 92 patients from the training cohort and tested in 72 patients from TCGA and TCIA. Encouragingly, the prediction model achieved AUC values of 0.79 in the training cohort and 0.77 in the TCGA testing cohort (Fig. 6f). This result reveals the possibility of noninvasive quantification for lncRNAs by deep learning radiomics.