Radiogenomic analysis of prediction HER2 status in breast cancer by linking ultrasound radiomic feature module with biological functions

The requirement for informed consent was waived because of the retrospective study design and the use of images and clinical information about patients derived from medical records. This study included patients who underwent preoperative ultrasound between January 2012 and December 2022 from the Second and the Third Affiliated Hospital of Harbin Medical University.

The patients were included when they met the inclusion criteria: a. pathologically confirmed breast cancer by core needle biopsy or surgical resection; b. ultrasound examination performed before invasive procedures; c. HER2 status of tumors was definite and d. evident lesions on ultrasound images. The exclusion criteria were listed as follows: a. patients who receive any treatment, such as neoadjuvant chemotherapy and radiotherapy, before ultrasound images collection; b. breast cancer of patients with equivocal HER2 status; c. patients with poor image quality; d. patients underwent invasive biopsy, which destroyed the image, before ultrasound images collection; e. patients with incomplete clinical data. Patient enrollment process of the study show in Additional file 1: Fig. S1.

The patient data comprised the following data sets: a. a region of interest (ROI) training set to develop a model that could automatically classify breast cancer/background; b. an ROI test set to evaluate the performance of the ROI classification model; and c. a radiomics analysis set that included breast cancer ultrasound images after accurate segmentation by the ROI classification model and radiologists to extract URFs; d. then we divided the patients from the radiomics analysis set into a radiomics training set and a test set for training and testing the model at a ratio of 7:3; e. finally, an independent validation set is added to further evaluate the performance of the model. The cohort selection flowchart is shown in Fig. 1. The flow diagram of study is shown in Fig. 2.

In this study, 489 breast cancer patients (mean age, 52 years ± 10 [standard deviation]; all women) were finally included. This study contained a total of 1859 ultrasound images, including 874 ultrasound images from 219 HER2-positive breast cancer patients and 985 ultrasound images from 270 HER2-negitive breast cancer patients. And five main data sets were included in this study for analysis: the ROI analysis set contained 49 patients, the ROI test set contained 345 patients, the radiomics training set contained 224 patients, the radiomics test set contained 96 patients and the independent validation set contained 95 patients.

All patients underwent breast cancer surgery or core needle biopsy to obtain tumor tissue for pathological diagnosis, right away after ultrasound images collection. The tumor tissues were stained with hematoxylin–eosin (HE) and performed in formalin-fixed, paraffin-embedded materials. After 5–7 days, the HER2 status of all breast cancer patients was tested using IHC or FISH according to the 2018 guideline recommendations of the American Society of Clinical Oncology and College of American Pathologists [30]. The IHC staining intensity of HER2 was scored as 0, 1 + , 2 + , or 3 + . The IHC staining intensity scored as 3 + was defined as HER2-positive, whereas the IHC staining intensity scored as 0 or 1 + was defined as HER2-negative. The IHC staining intensity scored as 2 + was defined as equivocal and FISH was used to further confirm HER2 status.

Ultrasound examinations were performed using HITACHI Vision 500 system (Hitachi Medical System, Tokyo, Japan) and Aixplorer ultrasound imaging system (SuperSonic Imagine, SSI, France), both equipped with a linear probe of 5–13 MHz. Radiologist with over five years of experience were selected for image collection and all radiologists were rigorous trained before collecting ultrasound images. They scanned the lesions from multiple angles, and selected several clear ultrasound images to be used for subsequent analysis.

We used Mask-R-Convolutional Neural Network-based architecture for the breast cancer/background classification of ultrasound images. We initially trained the convolutional layer in the Microsoft Common Objects in Context dataset with a learning rate of 0.001 and 1,000 epochs of 100 batches. Then, the deep learning (DL) algorithm was trained using the ROI training set and tested using the ROI test set.

The radiomic features in ROIs were extracted using the Python package, PyRadiomics (version 2.1.0; https://​pyradiomics.​readthedocs.​io/​) [31]. Eighty-six features were extracted from ROIs on the ultrasound images, and differences in features between HER2-positive and -negative samples were identified using Wilcoxon tests.

Three mRNA expression profiles [GSE45827, GSE129559 (median age, 55 years; range, 32–82), GSE162228] for HER2-positive breast cancer were downloaded from the GEO database based on the following inclusion criteria: (a) the organism was Homo sapiens; (b) HER2 status was explicit; (c) the patients were not previously treated. We ensured that the experiments were robust by using three datasets from different research institutions, sequencing platforms, and sequencing methods (high-throughput or array). Since features and genes are different dimensions of data, we applied the formula maximum value normalization as follows:where, Value_i is the feature or gene expression value of sample i, and max (Value_j) is the maximum value of the feature or gene j.

We identified positive-specific gene sets in 75 perturbation studies, each of which contained the same number of randomly selected samples from the URF sets and gene sets of HER2-positive samples. Pearson correlation coefficients (PCC) between the URFs and gene expression in the HER2-positive and HER2-negative samples were calculated to determine the positively- and negatively-related gene sets, respectively, corresponding to each URF. The negative URF-related genes were excluded from the positive URF-related gene sets to obtain positive-specific gene sets.

Functional enrichment of each URF in GO terms (p < 0.05) and KEGG pathways (p < 0.05) was analyzed using the R package “clusterProfiler” (Version 4.2.2; http://​www.​bioconductor.​org/​packages/​release/​bioc/​html/​clusterProfiler.​html) [32]. We obtained enriched GO terms and KEGG pathways in each URFs. Although some URFs might not be differential, they share similar functions as the differential URFs. We identified such URFs by calculating the Simpson index for each differential URFs. The formula is as follows:where F_n is the number of functions shared by features i and j, F_i is the number of functions in feature i, and F_j is the number of functions in feature j.

If the Simpson index of a pair URF was > 0.6, the URF was defined as an auxiliary URF, and the URF-module consisting of auxiliary and differential URFs was defined.

We downloaded mRNA expression profiles of breast cancer from GEO (GSE81538) as the validation set and normalized the maximum value. We performed 25 perturbation studies on the validation set, and positive-specific gene sets were identified in at least two perturbation studies. The functions associated with each URF in these positive-specific gene sets were validated using functional enrichment analysis.

We trained seven machine learning models (support vector machine, random forest, decision tree, logistic regression, Naive Bayes, artificial neural network, and K-nearest neighbor) based on eight differential URFs. We used 320 samples from the radiomics analysis set, 70% of it being the training set and 30% as the test set, and perturbed each model 10,000 times. The performance of the models was evaluated using receiver operator characteristics (ROC) curves.

To evaluate whether the URF-module could improve the performance of the classifiers, we retrained the seven models based on the URF-module and plotted ROC curves.

The CNN model we developed above was applied to segment 465 images from 95 patients in the independent validation set, and the segmentation results were identified by experts. Qualified ROIs are extracted by applying the Pyradiomics package for URFs, and the mean value was taken for multiple images of a patient. Finally, the performance of the differential URFs and URF-module in the independent validation set were verified in seven classifiers and ROC curves were plotted for evaluation.

The clinical characteristics of patients are summarized in Tables 1 and 2. To determine whether there were differences of clinical characteristics between HER2-negative and HER2-positive samples, we analyzed seven clinical characteristics (age, T, N and M status, lymph node status, TNM stage and menopausal status) in the test and validation set samples using the Wilcoxon nonparametric test. The results showed that there were no significant differences among the seven clinical characteristics, including menopause. These indicated that clinical characteristics did not effect the differential identification of radiomic features (p > 0.05).

The segmentation model initially identified the tumor region of the breast cancer samples (Fig. 3a), then, 989 of the 1,002 ultrasound images from 321 patients (radiomics analysis set) were accurately segmented with 98.7% accuracy (98.2% and 99.2% for the positive and negative sets, respectively). We then extracted 86 radiomics features (Fig. 3b) for the tumor region, which included First Order Statistics (18 features), Gray Level Co-occurrence Matrix (GLCM; 22 features), Gray Level Dependence Matrix (GLDM; 14 features), Gray Level Run Length Matrix (GLRLM; 16 features), and Gray Level Size Zone Matrix (GLSZM; 16 features).

We analyzed eight differential URFs (Size Zone Non Uniformity Normalized, Zone Entropy, Short Run High Gray Level Emphasis, Size Zone Non Uniformity, Small Area Emphasis, Short Run Emphasis, Run Length Non Uniformity Normalized, Small Area High Gray Level Emphasis) based on the fold-change values (fold change values > 1 and p < 0.05, respectively) between the positive and negative samples (Fig. 3c, d). These differential features contained GLSZM and GLRLM categories, indicating different patterns of features in these samples.

To examine whether radiomic features differed among different tumor locations and menopausal status, we tested the URFs of the train & test and validation set patients based on ANOVA (Additional file 1: Figs. S2 and S3). The results showed no significant differences in radiomic features among different tumor location and menopausal status (p > 0.05), further indicating that these clinical characteristics did not affect the differential identification of radiomic features.

The three breast cancer mRNA expression profiles of the test set included data from 324 patients (Fig. 4a). The distribution of positive and negative samples did not significantly differ (p > 0.05; chi-square test), indicating that it does not lead to biased results; however, the mRNA expression profiles between the two groups differed (Fig. 4b–d). We calculated correlations between differential URFs and genes in each dataset using random perturbation and Pearson correlation analysis (p < 0.05, |R|> 0.3). A gene was considered positive-specific only when it significantly associated with positive samples at least six times. We identified a minimum of 1,364 and a maximum of 1,871 positive genes in the eight differential URFs (Fig. 4e). Among the 25 genes with the most associations, we determined that a minimum of 8 and a maximum of 33 were significantly associated (p < 0.05). Notably, some URFs were associated with significantly more number of genes. For example, the glszm_Size Zone Non Uniformity(SZNU) feature was associated with 1,871 genes in 75 perturbations and ANKHD1 was significantly associated with all the differential URFs.

The results of the functional enrichment analysis of the positive-specific gene set (p < 0.05) showed that although the significance of differential URFs in the KEGG pathway was lower than that of the molecular function related GO terms, that of the ratio of genes was the highest (Fig. 5a). We then analyzed the functions of the differential URFs, and the results indicated that the Size Zone Non Uniformity Normalized (SZNUN) feature was mainly associated with nuclear division and intercellular communication, Zone Entropy (ZE) participated in immune cell activity and oxidative stress, Short Run High Gray Level Emphasis (SRHGLE) might contribute to cellular hypoxia, Size Zone Non Uniformity (SZNU) affected the catabolism of various compounds, small area emphasis(SAE) was associated with cellular hypoxia and protein metabolism, short run emphasis(SRE) regulated the cell cycle and affected neutrophil activity, Run Length Non Uniformity Normalized (RLNUN) was associated with the cell cycle and intercellular communication, and Small Area High Gray Level Emphasis (SAHGLE) was associated with endosomal localization and structure (Fig. 5b). We found a minimum of one and a maximum of six auxiliary features for each differential URF (Simpson index > 0. 6). Although most of the URFs were specific, some shared similar functions (Fig. 5c). We constructed a regulatory network for the URF-module (Fig. 5d), and the results confirmed known relationships and revealed novel relationships among these URFs.

The mRNA expression data, GSE81538, included 314 negative and 87 positive samples. To verify the functions of differential URFs, we identified a minimum of 62 and a maximum of 248 positive-specific genes among the eight differential URFs using Pearson correlation analysis and perturbation studies (p < 0.05, |R|> 0.3) (Fig. 6a). Among the 25 genes with the highest number of associations, we determined a minimum of 3 and a maximum of 4 significant associations. The results of analyzing the functional enrichment of the positive-specific gene sets (p.adjust < 0.05) showed that the most significant genes and the highest ratio of genes in KEGG pathways associated with the differential URFs and the second most significant genes in the GO terms were related to biological processes (Fig. 6b).

The functions of five differential URFs (fold change values > 1 and p < 0.05, respectively) were validated, and the results revealed that SZNU was associated with carbon metabolism in cancer, SZNUN was associated with cell-substrate junction, SAE is associated with chemical carcinogenesis - reactive oxygen species, SAHGLE was associated with the composition of the endomembrane, and ZE was involved in peroxidase and oxidoreductase activity (Fig. 6c).

To evaluate the ability of URFs to identify HER2 status, we first trained seven machine learning classifiers, including support vector machine, logistic, Bayes, Decision-Tree, random-forest, artificial neural network, and the K Nearest Neighbor algorithm based on the eight differential URFs (Fig. 7a–g). We then plotted ROC curves to facilitate and analyze comparisons among the different classifiers by measuring areas under ROC curves (AUC). An AUC of 1.0 indicated that the tested classifier was suitable for our data, whereas 0.5 indicated that the classifier had no classification capability for our data. The range of AUC of the tested classifiers was 0.62–0.715, suggesting that the logistic classifier achieved relative accurately classifying HER2-positive breast cancer patients.

We further evaluated the performance of the classifiers based on the URF-module. It can be found that the overall performance of the URF-module based model were better than those of URFs based model (Additional file 1: Table S1). In particular, the specificity of URF-module based model was 80.8% using logistic classifier in test set, implying a low rate of false positive. These results indicated that the URF-module were meaningful for identifying patients who are HER2-positive breast cancer and that the classification accuracy and specificity were improved (Fig. 7h).

To validate the ability of URFs to identify HER2-positive breast cancers, we evaluated the features extracted from the independent validation set based on the seven optimal classifiers. For each classifier, the AUC values ranged from 0.516 to 0.585. We also validated the performance of the URF-module based classifier with AUC values between 0.538 and 0.655(Fig. 8). Similar to test set, the overall performance of the URF-module based model were better than those of URFs based model (Additional file 1: Table S1). In particular, the specificity of URF-module based model was 83.3% using logistic classifier in independent validation set, also mean that a low rate of false positive. These results indicated that the URF-module can improve the classifier's ability to classify HER2-positive breast cancer patients.