Integrative Radiomics Clustering Analysis in Breast Cancer: Deciphering Heterogeneity and Prognostic Indicators through multiparametric MRI

doi:10.21203/rs.3.rs-3790296/v1

Download PDF

Article

Integrative Radiomics Clustering Analysis in Breast Cancer: Deciphering Heterogeneity and Prognostic Indicators through multiparametric MRI

https://doi.org/10.21203/rs.3.rs-3790296/v1

This work is licensed under a CC BY 4.0 License

Journal Publication

published 07 Aug, 2024

Read the published version in npj Breast Cancer →

You are reading this latest preprint version

Background

Breast cancer diagnosis and treatment have been revolutionized by advances in imaging techniques, particularly multiparametric Magnetic Resonance Imaging (mpMRI). This study aims to leverage mpMRI to enhance the understanding of breast cancer heterogeneity and improve diagnostic accuracy.

Methods

We conducted a comprehensive analysis of 194 breast cancer patients using mpMRI, which included T2-weighted imaging (T2WI), ZOOMit-Diffusion weighted imaging (ZOOMit-DWI), and Dynamic Contrast-Enhanced (DCE) MRI phase 2 and phase 7. Radiomics features were extracted using the open-source Python package 'pyradiomics'. Unsupervised analysis was performed using the MOVICS package, integrating various multi-omics clustering methods. The patients were clustered into different subtypes, and the associations between the subtypes and clinical prognostic indicators were investigated using univariate analyses.

Results

In total, 194 patients were included in the study with a mean age of 54.9 years and a predominance of Luminal B subtype (47.7%), followed by Luminal A (22.3%). Two mpMR-based radiomics subtypes were identified. Patient demographics showed no significant age difference between Subtype1 (average age 55.3 years, n = 109) and Subtype 2 (average age 54.4 years, n = 84) (p = 0.57). A significant disparity in molecular subtypes was observed between the groups, particularly in Luminal A subtype distribution (16.514% in Subtype 1 vs. 29.762% in Subtype 2, p = 0.03). Significant differences were also noted in hormone receptor status, with estrogen receptor (ER) (p = 0.01) and progesterone receptor (PR) (p = 0.04) differing notably between two subtypes. Group 1 presented a larger mean tumor size (p<0.01) and more varied histological grades (p<0.01). Lymph node metastasis (LNM) and edema showed significant differences (p<0.05).

Conclusions

Our study underscores the potential of mpMRI in enhancing breast cancer diagnostics. The radiomics-based cluster analysis offers a novel approach to categorizing breast cancer, providing insights into tumor heterogeneity and aiding in the development of personalized treatment strategies.

Biological sciences/Cancer/Breast cancer

Health sciences/Oncology/Cancer/Breast cancer

Breast cancer, a leading cause of cancer-related mortality among women worldwide, presents significant diagnostic challenges due to its heterogeneity^[1]. Recent advancements in multiparametric MRI have marked a significant leap in breast cancer diagnostics. This technique combines various imaging sequences to offer a comprehensive assessment of tumor characteristics, even capturing the complex heterogeneity within breast tumors^[2].Radiomics, an emerging field, further augments the potential of multiparametric MRI. It involves extracting a large number of quantitative features from medical images, uncovering patterns that may not be visible to the human eye^[3]. These features encompass a wide array of data points, including shape, intensity, texture, and wavelet features, each offering insights into the tumor's phenotype and underlying biology. The power of radiomics lies in its ability to transform static images into mineable data, opening new avenues for understanding tumor characteristics and their clinical implication.

Recent advancements in multiparametric MRI have marked a significant leap in breast cancer diagnostics. This technique, which combines various imaging sequences, offers a comprehensive assessment of tumor characteristics. Each sequence in multiparametric MRI contributes uniquely to the overall picture: T2-weighted imaging provides detailed anatomical information, diffusion-weighted imaging offers insights into cellular density, and dynamic contrast-enhanced sequences shed light on vascular properties of the tumor^[4]. Many studies underscore the potential of multiparametric MRI and radiomics in enhancing our understanding of breast cancer. However, they also highlight the complexity of this approach^[5]. The sheer volume and variety of data generated by radiomics analysis require sophisticated algorithms and machine learning techniques for interpretation. Lee et al.'s work on using machine learning for radiomic analysis is a testament to this, where they employed various algorithms to correlate MRI parameters with prognostic biomarkers and molecular subtypes of breast cancer^[6]. The integration of these advanced imaging and analytical techniques represents a significant stride towards personalized medicine in breast cancer care^[7]. It allows for a more nuanced understanding of the disease, potentially leading to more tailored and effective treatment strategies. However, the field is still evolving, and ongoing research is crucial to fully harness the potential of these technologies.

Despite these advancements, there remains a notable research gap in the specific application of radiomics clustering analysis in breast cancer using advanced MRI technologies^[8]. Clustering, an unsupervised learning technique, has effectively been applied to identify inherent groupings in patient samples, utilizing a specific similarity measure for visualization and partitioning purposes, the application value in the analysis of histopathological characteristics and prognosis of stage I lung adenocarcinoma has been reported^[9]. This study aims to bridge this gap by evaluating the application value of multiparametric imaging radiomics clustering analysis in breast cancer, utilizing the advanced capabilities of the 3.0T MRI scanner. The study focuses on a detailed analysis of radiomics features extracted from multiparametric MRI scans, correlating these features with histopathological data to enhance the accuracy and specificity of breast cancer classification. By doing so, the study endeavors to contribute to the nuanced understanding of breast cancer phenotypes and to aid in the development of more personalized and effective treatment strategies.

Study Sample

The study retrospectively analyzed breast cancer cases from MA’ANSHAN People's Hospital, spanning from September 2019 to October 2023. The age range of the subjects was between 32 and 85 years. Inclusion Criteria: (i)Complete MRI findings. (ii) No prior therapeutic interventions. (iii) Pathological results and immunohistochemical data within a week following the MRI examination. In this study, several exclusion criteria were applied. These included patients who had undergone prior puncture or radiotherapy (n = 3), those who received neoadjuvant chemotherapy before surgery (n = 4), and cases with incomplete MRI technique or poor image quality (n = 2). Additionally, patients were excluded if they lacked necessary immunohistochemical findings (n = 2), if the software algorithm produced incomplete values during the radiomic feature extraction process (n = 2) or The alignment was unsuccessful (n = 2). In all, 194 women were included in the study (Fig. 1). Ethical Considerations: The study was conducted following the ethical standards of the institutional and national research committee. It was approved by the Ethics Committee of MA’ANSHAN People’s Hospital (approval number: 2022007005) and the requirement for written informed consent was waived.

MRI Image Acquisition

Imaging was performed using the German Siemens Prisma 3.0T superconducting magnetic resonance scanner with an 18-channel phased array breast-specific coil. Patients were positioned prone, feet first. The imaging techniques included T2-weighted imaging (T2WI), ZOOMit-Diffusion weighted imaging (ZOOMit-DWI), and Dynamic Contrast-Enhanced (DCE) MRI phases 2 and 7.

Axial T1 Turbo Spin Echo (TSE) Sequence:Field of View (FOV): 340×340 mm, Matrix: 384×286, Repetition Time (TR): 5.3 ms, Echo Time (TE): 2.46 ms, Slice Thickness: 1.5 mm, Axial T2 TSE Sequence: FOV: 340×340 mm, Matrix: 384×286, TR: 6000 ms, TE: 63 ms, Slice Thickness: 4 mm, Interval: 0.8 mm, ZOOMit-DWI Sequence: Axial Scan Sequence, FOV: 340×340 mm, Matrix: 190×190, TR: 6000 ms, TE: 63 ms, Slice Thickness: 4.0 mm, Interval: 0.8 mm, b-values: 50, 500, 1000, 1500 s/mm², Dynamic Contrast-Enhanced MRI (DCE-MRI) Sequence: FOV: 360×360 mm, Matrix: 384×286, TR: 4.13 ms, TE: 1.3 ms, Slice Thickness: 1.2 mm, High-resolution T1-weighted three-dimensional fast gradient-echo sequence, Consisting of nine phases, each with a scanning time of 60 seconds, The contrast agent (Omniscan, GE Pharmaceuticals) was administered at the end of the first phase, Subsequent phases had scanning times of 60s, 120s, 240s, 300s, 360s, 420s, 480s, and 540s, Enhanced scanning was performed using a high-pressure injector, Contrast agent dosage: 0.1 mmol/kg of body weight, flow rate of 2.5 mL/s, Followed by a 20 mL saline flush(Table 1).

Table 1

Imaging parameters of breast MRI
Parameter	T1WI	T2WI	DWI	DCE-MRI
Sequence Type	TSE	TSE	ZOOMit-DWI	Fast Gradient-Echo
Field of View (mm)	340×340	340×340	340×340	360×360
Matrix	384×286	384×286	190×190	384×286
TR (ms)	5.3	6000	6000	4.13
TE (ms)	2.46	63	63	1.3
Slice Thickness(mm)	1.5	4	4	1.2
Interval (mm)	Flash-3D	0.8	0.8	Flash-3D
b-value (s/mm²)	-	-	50, 500, 1000,1500	-
Contrast Agent (DCE-MRI only)	-	-	-	Omniscan (GE Pharmaceuticals)
Injection Rate (mL/s)	-	-	-	2.5
Phases (Scanning and Enhancement)	-	-	-	1 + 8 phases
Scanning Time per Phase (s)	-	-	-	60s
Abbreviations: TSE, Turbo Spin Echo; TR, Repetition Time; TE, Echo Time.

Table 2

Clinical, Histologic, and Radiological characteristics of patients
Characteristic	Level	Finding (n = 194)
Age (years)*	Mean ± SD	54.9 ± 10.8
Molecular subtypes	Luminal A	43 (22.3)
	Luminal B	92 (47.7)
	HER2-positive	18 (9.3)
	Basal-like	40 (20.7)
ER	Negative	56 (29.0)
	Positive	137 (71.0)
PR	Negative	98 (50.8)
	Positive	95 (49.2)
Her_2	Negative	80 (41.5)
	Positive	113 (58.5)
Ki67	<14%	69 (35.8)
	≥ 14%	124 (64.2)
Maximum(cm)	Mean (SD)	2.51 (1.53)
Histological grades	I	55 (28.5)
	II	117 (60.6)
	III	21 (10.9)
LNM	Negative	113 (58.5)
	Positive	80 (41.5)
LVI	Negative	143 (74.1)
	Positive	50 (25.9)
Around	Left	93 (48.2)
	Right	100 (51.8)
FGT	a	9 (4.7)
	b	60 (31.1)
	c	100 (51.8)
	d	24 (12.4)
Location	Superior external	87 (45.1)
	Inferior external	28 (14.5)
	Superior internal	40 (20.7)
	Inferior internal	16 (8.3)
	Posterior areola	14 (7.3)
	Central region	8 (4.1)
BEP	Minimal	72 (37.3)
	Mild	98 (50.8)
	Moderate	20 (10.4)
	Significant	3 (1.6)
Type	Non-mass	17 (8.8)
	Mass	176 (91.2)
TIC	I	7 (3.6)
	II	74 (38.3)
	III	112 (58.0)
Edema	No	95 (49.2)
	Yes	98 (50.8)
T2WI signals	Low	50 (25.9)
	Equal	72 (37.3)
	High	71 (36.8)
DWI signals	Low	1 (0.5)
	Equal	13 (6.7)
	High	179 (92.7)
Abbreviations: ER, estrogen receptor; PR, progesterone receptor; LNM, lymph node metastasis; LVI, lymphovascular invasion; FGT, fibroglandular tissue; BEP, background parenchymal Enhancement; TIC,time signal intensity curve; SD, standard deviation. Unless otherwise indicated, data are numbers of patients and data in parentheses are percentages. * Data in parentheses are mean ± SD.

Table 3

Clinical, Histologic, and Radiological characteristics of patients by Clustering
Characteristic	Level	Group = 1 (n = 109)	Group = 2 (n = 84)	p-value
Age (years)*	Mean ± SD	55.3 ± 10.6	54.4 ± 11.0	0.574
Molecular subtypes	Luminal A	18 (16.514)	25 (29.762)	0.034
	Luminal B	50 (45.872)	42 (50.000)
	HER2-positive	13 (11.927)	5 (5.952)
	Basal-like	28 (25.688)	12 (14.286)
ER	Negative	40 (36.697)	16 (19.048)	0.012
	Positive	69 (63.303)	68 (80.952)	0.012
PR	Negative	63 (57.798)	35 (41.667)	0.038
	Positive	46 (42.202)	49 (58.333)	0.038
Her_2	Negative	44 (40.367)	36 (42.857)	1
	Positive	65 (59.633)	48 (57.143)	1
Ki67	<14%	34 (31.193)	35 (41.667)	0.176
	≥ 14%	75 (68.807)	49 (58.333)	0.176
Maximum(cm)	Mean (SD)	3.184 (1.678)	1.643 (0.613)	< 0.01
Histological grades	I	18 (16.514)	37 (44.048)	< 0.01
	II	78 (71.560)	39 (46.429)
	III	13 (11.927)	8 (9.524)
LNM	Negative	55 (50.459)	58 (69.048)	0.014
	Positive	54 (49.541)	26 (30.952)	0.014
LVI	Negative	76 (69.725)	67 (79.762)	0.158
	Positive	33 (30.275)	17 (20.238)	0.158
Around	Left	49 (44.954)	44 (52.381)	0.380
	Right	60 (55.046)	40 (47.619)	0.380
FGT	a	3 (2.752)	6 (7.143)	0.272
	b	34 (31.193)	26 (30.952)
	c	61 (55.963)	39 (46.429)
	d	11 (10.092)	13 (15.476)
Location	Superior external	43 (39.450)	44 (52.381)	0.022
	Inferior external	16 (14.679)	12 (14.286)
	Superior internal	21 (19.266)	19 (22.619)
	Inferior internal	10 (9.174)	6 (7.143)
	Posterior areola	14 (12.844)	0 (0.000)
	Central region	5 (4.587)	3 (3.571)
BEP	Minimal	39 (35.780)	33 (39.286)	0.126
	Mild	52 (47.706)	46 (54.762)
	Moderate	15 (13.761)	5 (5.952)
	Significant	3 (2.752)	0 (0.000)
Type	Non-mass	10 (9.174)	7 (8.333)	1
	Mass	99 (90.826)	77 (91.667)	1
TIC	I	3 (2.752)	4 (4.762)	0.0726
	II	35 (32.110)	39 (46.429)
	III	71 (65.138)	41 (48.810)
Edema	No	33 (30.275)	62 (73.810)	< 0.01
	Yes	76 (69.725)	22 (26.190)	< 0.01
T2WI signals	Low	24 (22.018)	26 (30.952)	0.283
	Equal	45 (41.284)	27 (32.143)
	High	40 (36.697)	31 (36.905)
DWI signals	Low	1 (0.917)	0 (0.000)	0.508
	Equal	6 (5.505)	7 (8.333)
	High	102 (93.578)	77 (91.667)
Abbreviations: ER, estrogen receptor; PR, progesterone receptor; LNM, lymph node metastasis; LVI, lymphovascular invasion; FGT, fibroglandular tissue; BEP, background parenchymal Enhancement; TIC,time signal intensity curve; SD, standard deviation. Bold indicates statistical significance. Unless otherwise indicated, data are numbers of patients and data in parentheses are percentages. * Data in parentheses are mean ± SD.

Image Processing

The ADC image was specifically generated based on a b-value of 1500 s/mm². This parameter was chosen as it is typically optimal for highlighting differences in tissue diffusivity, which is particularly relevant in the context of breast tumor characterization^[4].

For the assessment of tumor morphology, all MRI images underwent a thorough review by two experienced breast radiologists, who had 10 and 23 years of experience, respectively, in the diagnostic imaging of breast diseases. This review was conducted while the radiologists were blinded to the patients' clinical history. The evaluation adhered to the criteria set forth in the 2013 Breast Imaging Reporting and Data System (BI-RADS) MR lexicon^[10]. In cases where patients presented with multiple lesions, the largest lesion was selected for detailed evaluation. The radiologists recorded several key aspects of each lesion, including: Location and Type (mass or non-mass), morphology (round, round-like, or irregular), Margin (clear or indistinct), Maximum Diameter (measuring the largest dimension of the lesion), Fibroglandular Tissue (FGT) and Background Parenchymal Enhancement (BPE). Additionally, Signal Intensity on T2WI and DWI (slightly low, equal, or high signal) and edema in Adjacent Breast Parenchyma.

The presence of intratumoral high signal intensity (SI) was meticulously assessed on T2WI and DWI sequences, and peritumoral edema was meticulously assessed on T2WI^{[11, 12]}. Intratumoral high SI was visually defined when the lesion's SI was stronger than that of vessels, water, or the surrounding parenchymal tissue, and peritumoral edema was determined based on the observation of high SI around the tumor on T2WI^{[11, 12]}.

Tumor Segmentation

The tumor boundary was delineated by two experienced radiologists (Radiologist 1, Q.X. 10 years in breast cancer diagnosis. Radiologist 2, H.Y.S. 23 years in breast cancer diagnosis.) using ITK-SNAP software (http://www.itk-snap.org). Q.X. carefully segmented the tumor on the enhanced phase 2 images, including cystic and necrotic areas. H.Y.S. checked the segmentation, if they had different opinions, they discussed to reach a consensus. After segmentation, a three-dimensional Volume of Interest (VOI) were obtained (Fig. 2). To ensure consistency across the different modalities, we used the “Registration” module in SPM software to align images to the second phase image. This alignment was crucial for accurately matching the Volume of Interest (VOI) derived from the second phase with corresponding areas in other modalities. After alignment, the VOI and images were imported into ITK-SNAP for a thorough visual inspection. Any discrepancies identified led to manual adjustments of the VOI, ensuring precise and uniform tumor segmentation across the different MRI sequences.

Radiomics Feature Extraction and Consensus Clustering

Radiomics feature extraction for each sequence of mpMRI was conducted using the open-source Python package 'pyradiomics' (version 3.0.1, https://pyradiomics.readthedocs.io/en/latest/), an IBSI-compliant tool. The feature extraction included two steps, image preprocessing and feature calculation. The preprocessing of images included resampling image to an isotropic voxel size of 3mm and grayscale discretization with a bin width of 10. Additionally, image intensity normalization was applied. In total, 1158 radiomics features were extracted for each modalities, encompassing shape, first-order, and second-order texture features, as well as high-order features derived from wavelet-transformed and Laplace of Gaussian transformed images.

In our study, we employed unsupervised clustering analysis to categorize patients using the multi-parametric’s radiomics features, including T2WI, ADC, DCEphase2, and DCEphase7. This comprehensive clustering analysis was conducted on the entire cohort of 194 patients, ensuring a robust and thorough exploration of the radiomics data and its potential clinical correlations. The process involved several key steps: (1) Feature Dimension Reduction: Initially, we applied the Uniform Manifold Approximation and Projection (UMAP) method for feature dimension reduction^[13]. This step was crucial in simplifying the complex radiomics data while retaining essential information. For each imaging modality, we preserved 10 key features, ensuring a comprehensive yet manageable dataset for subsequent analysis. (2) Determination of the optimal number of clusters: The cluster prediction index (CPI) was used to identify the optimal number of clusters (k), which were calculated at the range of k from 2 to 8, and the number of clusters with maximum CPI was determined as the optimal number of clusters. (3) Integrative clustering: For our data was multi-modalities, similar to multi-omics data, so we applied the advanced multi-omics clustering algorithms from the MOVICS package to perform multi-modalities clustering analysis^[14]. To obtain the robust clustering results, 10 advanced multi-omics clustering analysis methods were performed, and the consensus matrix counting the times of paired patients in the same cluster was obtained and determined which cluster the patients belonged to. (4) Comparison of Radiomics features in different clusters: Once the clusters were established, the radiomics feature distribution heatmap was given to show whether the features’ distributions were different between different cluster.

Histological Evaluation

Lymph node metastasis (LNM) occurrence, lymphovascular invasion (LVI) status, and pathological type were recorded post surgery. LVI was positive when cancer cells were found in lymphatic vessels or small blood vessels outside the tumor. Streptavidin-peroxidase immunohistochemistry was used to determine the expression statuses of estrogen receptor (ER), progesterone receptor (PR), HER2, and Ki-67 in surgically excised breast cancer tissues. Molecular subtypes were categorized based on these markers: Luminal A (ER/PR ≥ 1%, HER2-negative, Ki-67 < 14%), Luminal B (ER/PR ≥ 1% with either Ki-67 ≥ 14% or HER2-positive), Basal-like (ER < 1%, PR < 1%, HER2-negative), and HER2-positive (ER < 1%, PR < 1%, HER2-positive). Ki-67 expression was categorized as low (< 14%) or high (≥ 14%). Histologic grade, determined using the Nottingham grading system, was categorized as low malignancy (grade I, score 3–5), moderately malignant (grade II, score 6–7), and highly malignant (grade III, score 8–9), with the World Health Organization adopting this system as the standard for invasive breast cancer. ER or PR positivity was defined as positive nuclear staining in more than 10% of cells, while HER2 positivity was indicated by a CerbB-2 value of 3 + or a fluorescence in situ hybridization (FISH) amplification ratio ≥ 2.0.

Statistical Analysis

All statistical analyses were conducted using R software (version 4.2.1, www.r-project.org). Continuous variables were described as mean ± standard deviation when normally distributed, and as median with interquartile range (IQR) for non-normal distributions. To compare clinical characteristics across the final clusters, categorical data were analyzed using the Fisher’s exact test or the chi-square (χ²) test, while continuous data were evaluated using the Kruskal-Wallis test. All tests were two-sided, with a P-value of less than 0.05 considered statistically significant.

DATA AVAILABILITY

Data generated in this study are available upon reasonable request from the corresponding author.

CODE AVAILABILITY

The code used for statistical calculations in this analysis is available upon reasonable request from the corresponding author.

Patient and Tumor Characteristics

In a study of 194 breast cancer patients, key findings include a mean age of 54.9 years and a predominance of Luminal B subtype (47.7%), followed by Luminal A (22.3%). Hormone receptor analysis showed 71.0% ER-positive and 49.2% PR-positive patients, with 58.5% exhibiting HER2 positivity. Ki67 expression was high in 64.2% of patients. The average tumor size was 2.51 cm. Histologically, most tumors were Grade II (60.6%). Lymph node metastasis was present in 41.5% of cases, and lymphovascular invasion in 25.9%. Fibroglandular tissue distribution varied, with the majority in categories c (51.8%) and b (31.1%). Background parenchymal enhancement was mostly minimal (37.3%) or mild (50.8%). The majority of tumors were mass type (91.2%), with TIC Type III being the most common (58.0%). Edema was present in 50.8% of cases, and DWI signals were predominantly high (92.7%) (Table 1). These comprehensive results provide a detailed overview of the clinical and pathological characteristics of the breast cancer cases in this study.

Clustering Analysis

The data indicates that the Cluster Prediction Index (CPI) reaches its peak at k = 2, suggesting that a classification count of 2 is the most optimal for this analysis(Fig. 3). Setting k = 2, ten advanced multi-omics clustering analysis methods are employed for clustering. A consistency analysis of these ten results yields a clustering consistency matrix(Fig. 4) and a heatmap of the distribution patterns of radiomic features in two different groups. It clearly shows that the radiomic features vary between the different groups(Fig. 5).

Consensus Clustering Associations

Patient Clinical Characteristics

Begin by presenting the basic demographic information: The mean age of patients in Group 1 (n = 109) was 55.3 ± 10.6 years, and in Group 2 (n = 84), it was 54.4 ± 11.0 years. The difference in age between the two groups was not statistically significant (p = 0.57).

Analysis of Breast Cancer Subtypes and Receptor Status

In the results, it was observed that the distribution of molecular subtypes across both groups showed a significant difference (p = 0.03). In our study, the distribution of the molecular subtype Luminal A differed notably between the two groups, with 16.514% in Group 1 and 29.762% in Group 2. This pattern is indicative of a significant variation in hormone receptor status across the groups. The differences were particularly pronounced in both ER and PR. A statistically significant difference was found in ER status between the groups(p = 0.01). Similarly, PR status also showed a significant variance (p = 0.04). These findings suggest distinct hormonal profiles within the respective cohorts, underlining the importance of hormonal characterization in such studies.

Tumor Characteristics

The maximum tumor size showed a significant difference, with Group 1 having a mean size of 3.184 ± 1.678 cm and Group 2 having 1.643 ± 0.613 cm (p<0.01). Histological grades also varied significantly between the groups (p<0.01), suggesting differences in tumor aggressiveness or stage.

Lymph Node and Other Pathological Features:

A significant difference was found in LNM (p<0.05), but not in LVI (p = 0.158). No significant differences were observed in parameters like Around, FGT, Location, BEP, Type, TIC, T2WI signals, and DWI signals, except for Edema, which showed a significant difference (p<0.01).

Clustering primarily serves as an analytical technique that extracts meaningful information based on the pattern characteristics of a target data set. It divides unlabelled data into different groups, where the similarity within each group is high, and the dissimilarity between groups is significant. Currently, this approach is applied in many fields^[15]. Our study adopts a hierarchical clustering method, providing a structured approach to discern intrinsic imaging phenotypes of breast tumors. Unlike previous research primarily reliant on qualitative interpretation for identifying distinct imaging patterns linked to histopathologic correlates^[16], our study introduces a systematic, quantitative evaluation framework. Cluster analysis aims to categorize data points within a population into groups where members of each cluster share more similarities with each other than with those in different clusters^[9]. We employed unsupervised cluster analysis to uncover natural groupings within the imaging features we extracted, leading to the identification of intrinsic phenotypes characterized by distinct and measurable imaging patterns. This method reflects the clustering techniques pivotal in uncovering the inherent molecular subtypes of breast cancer. To visualize and interpret these imaging phenotypes, we generated heat maps akin to microarray expression clustering dendrograms. This visualization aids in beginning to understand the correlation between the imaging presentation of cancer and its molecular profile, potentially offering insights into the likelihood of cancer recurrence. The study's findings reveal notable differences in several critical parameters between the two patient groups. The significant variations observed in these areas underscore the efficacy of radiomics clustering analysis in distinguishing breast cancer cases based on their intrinsic phenotypes. This differentiation is crucial, as it paves the way for more personalized treatment approaches, tailoring strategies to the specific characteristics and needs of each tumor type. The ability to categorize breast cancer more precisely based on these key parameters holds significant potential for enhancing the effectiveness of treatment plans and improving patient outcomes.

The cluster analysis results of our study, showing statistical significance in molecular subtypes, ER, PR, size, histological grading, edema, LNM, and location, align with recent advancements in breast cancer research. Lyu et al analyzed factors influencing sentinel lymph node metastasis in breast cancer, highlighting the importance of molecular subtypes, tumor size, and histological grade in predicting lymph node involvement^[17]. Wang et al analyzed risk factors of axillary lymph node metastasis and prognosis in T1 breast cancer using the SEER database, highlighting the impact of tumor size, histological grade, and subtype on lymph node involvement and survival outcomes^[18]. This further validates our study's emphasis on these factors in breast cancer assessment.This supports our findings on the significance of these parameters in breast cancer characterization. Xu et al investigated molecular phenotypes and clinical characteristics of familial hereditary breast cancer, emphasizing the role of clinical pathologic subtypes, tumor location, histological grade, and lymph node metastasis. The study results showed significant differencesbetween hereditary and non-hereditary breast cancers, including histological grading (grade II/III), LNM, PR, and HER2^[19]. Their findings resonate with our study, underscoring the complexity and diversity of breast cancer subtypes. Our study contributes significantly to breast cancer research by providing a comprehensive cluster analysis that correlates imaging features with key clinical and pathological parameters. We offer a novel approach to breast cancer classification, integrating imaging phenotypes with molecular typing, ER, PR, tumor size, histological grading, edema, and lymph node metastasis , these may have varying influences on the prognosis and survival outcomes in breast cancer patients. This multidimensional analysis enhances the understanding of breast cancer heterogeneity and supports the development of personalized treatment strategies.

Molecular subtypes, particularly those defined by hormone receptor status, have been widely recognized for their prognostic implications. ER-positive and PR-positive tumors often suggest a better prognosis and responsiveness to hormonal therapies, which could lead to improved survival rates^[20]. Conversely, tumors with negative hormone receptor status, particularly those falling into the triple-negative category, are often associated with a more aggressive course and poorer outcomes ^[21]. Lee et al demonstrated the potential of machine learning in radiomics for predicting prognostic biomarkers and molecular subtypes of breast cancer, emphasizing the role of tumor heterogeneity and angiogenesis properties on MRI^[6]. This aligns with our observation of distinct phenotypic profiles in breast cancer, highlighting the importance of personalized treatment strategies. Our results also align with Romeo et al, who applied a machine learning-based radiomics approach to hybrid 18F-FDG PET/MRI for predicting axillary lymph node involvement in breast cancer^[22]. This supports our findings on the utility of radiomics in assessing tumor aggressiveness and invasion potential. Furthermore, Araz et al explored the role of preoperative 18F-FDG PET/CT radiomics features in predicting hormone receptor positivity in primary breast tumors^[23]. While their study found limited predictive value in radiomics parameters for hormone receptor status, it highlights the complexity and challenges in correlating radiomics features with specific tumor characteristics, a theme also evident in our study. Metzger Filho et al assessed the impact of HER2 heterogeneity on response to HER2-targeted therapy, revealing that HER2 heterogeneity is associated with resistance to such treatments^[24]. Szep et al demonstrated that whole-tumor ADC texture analysis could predict hormonal status in breast cancer masses^[25]. These findings are particularly relevant to our study, as they highlighted the importance of understanding tumor heterogeneity in treatment planning, and emphasized on the potential of radiomics in breast cancer diagnosis and treatment.

Tumor size, histological grading, and tumor location are also critical in determining the aggressiveness of the disease. Larger tumors and those with higher histological grades are generally indicative of a more advanced disease state, potentially leading to lower survival rates. Min Y et al. identified clinicopathological characteristics associated with the risk of distant metastasis and prognosis in breast cancer patients without lymph node metastasis, highlighting the importance of tumor size and histological grade^[26]. Similarly, the presence of lymph node metastasis is a well-established marker of poor prognosis, as it signifies the spread of cancer cells beyond the primary tumor site. Lymph node metastasis in breast cancer patients is a critical factor that significantly influences prognosis. The presence of cancer cells in the lymph nodes, particularly the axillary lymph nodes, is often considered a hallmark of the spread of the disease beyond the primary tumor site. This metastatic spread is a key determinant in staging breast cancer, which in turn guides treatment decisions and helps predict patient outcomes. Patients with lymph node involvement generally have a poorer prognosis compared to those without. Recent studies, including the development of nomogram models, have highlighted several independent predictive factors for axillary lymph node metastasis in breast cancer patients, and these factors include tumor size, primary site, molecular subtype, and histological grading^[27].

Edema, while not as extensively studied as other factors, could provide additional insights into the tumor microenvironment and its interaction with the surrounding tissues, which might have implications for disease progression and metastatic potential. In the study by Zeyan Xu et al., edema is a promising predictive factor for lymph node metastasis in breast cancer, which is very important for patients with early-stage breast cancer and may help in formulating treatment plans. Additionally, a positive correlation was observed between BES and various aggressive clinicopathological factors (p < 0.05)^[28]. In future studies, it would be crucial to integrate these factors into survival analysis models to comprehensively understand their collective impact on patient outcomes. Such analyses could not only validate the clinical significance of our cluster analysis results but also contribute to the development of more tailored and effective treatment strategies, ultimately improving patient survival and quality of life in breast cancer.

Our study makes significant contributions to the field of breast cancer research by integrating radiomics clustering analysis with traditional histopathological markers. We provide a comprehensive approach to breast cancer classification, enhancing the understanding of tumor heterogeneity and its implications for personalized medicine. Our findings contribute to the growing body of evidence supporting the use of radiomics in breast cancer diagnosis and treatment planning. In evaluating conflicting explanations of our results, we defend our approach by highlighting the robustness of our methodology and the consistency of our findings with current literature. While some studies suggest limited predictive value of radiom^[28]ics in certain aspects of breast cancer, our study demonstrates its utility in differentiating intrinsic phenotypes, supporting its role in personalized treatment strategies.

While our study provides valuable insights into the classification and treatment of breast cancer using radiomics, it is important to acknowledge its limitations. Firstly, the limited sample size of 194 patients, although substantial, may not fully capture the diversity and complexity of breast cancer phenotypes. Another significant limitation is the absence of survival analysis. Our study focused on the classification of breast cancer based on radiomics and traditional histopathological markers, without delving into the long-term outcomes of these patients. Future studies with larger cohorts are necessary to validate and extend our results, and include survival data to provide a more comprehensive understanding of the clinical significance of radiomics-based classification. Additionally, our study did not incorporate genetic testing results, which could provide deeper insights into the molecular underpinnings of the observed phenotypes. The lack of genetic information represents a missed opportunity to explore the interplay between genotypic and phenotypic characteristics of breast cancer. In conclusion, while our study makes significant contributions to the field of breast cancer research, addressing these limitations in future studies will be crucial for advancing our understanding of breast cancer and improving patient care. Expanding the sample size, including survival analysis, and integrating genetic testing results are key areas for future research that could significantly enhance the impact and applicability of our findings.

FUNDING: This article is supported by the Key Research and Development Plan of Anhui Province in 2022 (No:2022e07020065); Ma'anshan City Science and Technology Planning Project (No:YL-2022-5); Ma'anshan City Health Commission Key Research Project (No:MASWJ2022a001).

AUTHOR CONTRIBUTIONS: Y.H.S., S.F.D. and W.L.W. provided equal contributions to the study. H.K.Y. conceptualized the study. S.Y.P., W.Q.C. and L.X. performed all radiologic analyses. X.Q. wrote the original draft of the manuscript and provided clinical supervision. All authors contributed to the study conception and design, material preparation, data collection and analysis.

COMPETING INTERESTS: The authors declare no competing interests

Turner KM, Yeo SK, Holm TM, Shaughnessy E, Guan JL. Heterogeneity within molecular subtypes of breast cancer[J]. Am J Physiol Cell Physiol,321(2): C343-C354 (2021).
Chhetri A, Li X, Rispoli JV. Current and Emerging Magnetic Resonance-Based Techniques for Breast Cancer[J]. Front Med (Lausanne), 7: 175 (2020).
Mayerhoefer ME, et al. Introduction to Radiomics[J]. J Nucl Med, 61(4): 488–495 (2020).
Daimiel Naranjo I, et al. Diagnostic value of diffusion-weighted imaging with synthetic b-values in breast tumors: comparison with dynamic contrast-enhanced and multiparametric MRI[J]. Eur Radiol, 31(1): 356–367 (2021).
Fan M, et al. Radiomics of Tumor Heterogeneity in Longitudinal Dynamic Contrast-Enhanced Magnetic Resonance Imaging for Predicting Response to Neoadjuvant Chemotherapy in Breast Cancer[J]. Front Mol Biosci, 8: 622219 (2021).
Lee JY, et al. Radiomic machine learning for predicting prognostic biomarkers and molecular subtypes of breast cancer using tumor heterogeneity and angiogenesis properties on MRI[J]. Eur Radiol,32(1): 650–660 (2022).
Umutlu L, et al. Multiparametric Integrated (18)F-FDG PET/MRI-Based Radiomics for Breast Cancer Phenotyping and Tumor Decoding[J]. Cancers (Basel), 13(12) (2021).
Fan M, et al. Multiparametric MRI radiomics fusion for predicting the response and shrinkage pattern to neoadjuvant chemotherapy in breast cancer[J]. Front Oncol, 13: 1057841 (2023).
Perez-Johnston R, et al. CT-based Radiogenomic Analysis of Clinical Stage I Lung Adenocarcinoma with Histopathologic Features and Oncologic Outcomes[J]. Radiology, 303(3): 664–672 (2022).
Vinnicombe S. How I report breast magnetic resonance imaging studies for breast cancer staging and screening[J]. Cancer Imaging, 16(1): 17 (2016).
Baltzer PA, et al. Sensitivity and specificity of unilateral edema on T2w-TSE sequences in MR-Mammography considering 974 histologically verified lesions[J]. Breast J, 16(3): 233–239 (2010).
Uematsu T, Kasami M, Watanabe J. Is evaluation of the presence of prepectoral edema on T2-weighted with fat-suppression 3 T breast MRI a simple and readily available noninvasive technique for estimation of prognosis in patients with breast cancer?[J]. Breast Cancer, 21(6): 684–692 (2014).
Mcinnes L, Healy J, Melville J. UMAP: Uniform Manifold Approximation and Projection for Dimension Reduction[Z]. Ithaca: Cornell University Library, arXiv.org (2020).
Lu X, Meng J, Zhou Y, Jiang L, Yan F. MOVICS: an R package for multi-omics integration and visualization in cancer subtyping[J]. Bioinformatics, 36(22–23): 5539–5541 (2021).
Oyewole GJ, Thopil GA. Data clustering: application and trends[J]. Artif Intell Rev, 56(7): 6439–6475 (2023).
Witowski J, et al. Improving breast cancer diagnostics with deep learning for MRI[J]. Sci Transl Med, 14(664): o4802 (2022).
Lyu W, Guo Y, Peng H, Xie N, Gao H. Analysis of the Influencing Factors of Sentinel Lymph Node Metastasis in Breast Cancer[J]. Evid Based Complement Alternat Med, 2022: 5775971 (2022).
Wang WY, et al. [Analysis of risk factors of axillary lymph node metastasis and prognosis in T1 breast cancer: a large-scale retrospective study based on the SEER database][J]. Zhonghua Yi Xue Za Zhi, 101(27): 2152–2158 (2021).
Xu Y, He J, Qian C, Yang C. Molecular phenotypes and clinical characterization of familial hereditary breast cancer among half and full sisters[J]. BMC Womens Health, 22(1): 145 (2022).
Hanker AB, Sudhan DR, Arteaga CL. Overcoming Endocrine Resistance in Breast Cancer[J]. Cancer Cell, 37(4): 496–513 (2020).
Zhao H, Gong Y. The Prognosis of Single Hormone Receptor-Positive Breast Cancer Stratified by HER2 Status[J]. Front Oncol, 11: 643956 (2021).
Romeo V, et al. Simultaneous 18F-FDG PET/MRI Radiomics and Machine Learning Analysis of the Primary Breast Tumor for the Preoperative Prediction of Axillary Lymph Node Status in Breast Cancer[J]. Cancers (Basel), 15(20) (2023).
Araz M, et al. Can Radiomics Analyses in (18)F-FDG PET/CT Images of Primary Breast Carcinoma Predict Hormone Receptor Status?[J]. Mol Imaging Radionucl Ther, 31(1): 49–56 (2022).
Filho OM, et al. Impact of HER2 Heterogeneity on Treatment Response of Early-Stage HER2-Positive Breast Cancer: Phase II Neoadjuvant Clinical Trial of T-DM1 Combined with Pertuzumab[J]. Cancer Discov, 11(10): 2474–2487 (2021).
Szep M, et al. Whole-Tumor ADC Texture Analysis Is Able to Predict Breast Cancer Receptor Status[J]. Diagnostics (Basel), 13(8) (2023).
Min Y, et al. Risk Factors, Prognostic Factors, and Nomogram for Distant Metastasis in Breast Cancer Patients Without Lymph Node Metastasis[J]. Front Endocrinol (Lausanne), 12: 771226 (2021).
Gao X, Luo W, He L, Yang L. Nomogram models for stratified prediction of axillary lymph node metastasis in breast cancer patients (cN0)[J]. Front Endocrinol (Lausanne), 13: 967062 (2022).
Xu Z, et al. MRI characteristics of breast edema for assessing axillary lymph node burden in early-stage breast cancer: a retrospective bicentric study[J]. Eur Radiol, 32(12): 8213–8225 (2022).

(Not answered)

Download PDF

Journal Publication

published 07 Aug, 2024

Read the published version in npj Breast Cancer →

Editorial decision: revise
06 Feb, 2024
Review #1 received at journal
04 Feb, 2024
Review #2 received at journal
02 Feb, 2024
Reviewer #2 agreed at journal
24 Jan, 2024
Reviewer #1 agreed at journal
16 Jan, 2024
Reviewers invited by journal
11 Jan, 2024
Editor assigned by journal
26 Dec, 2023
Submission checks completed at journal
22 Dec, 2023
First submitted to journal
22 Dec, 2023

You are reading this latest preprint version

Integrative Radiomics Clustering Analysis in Breast Cancer: Deciphering Heterogeneity and Prognostic Indicators through multiparametric MRI

Status:

Journal Publication

Version 1

Abstract

Figures

INTRODUCTION

MATERIALS AND METHODS

Study Sample

MRI Image Acquisition

Image Processing

Tumor Segmentation

Radiomics Feature Extraction and Consensus Clustering

Histological Evaluation

Statistical Analysis

DATA AVAILABILITY

CODE AVAILABILITY

RESULTS

Patient and Tumor Characteristics

Clustering Analysis

Consensus Clustering Associations

Patient Clinical Characteristics

Analysis of Breast Cancer Subtypes and Receptor Status

Tumor Characteristics

Lymph Node and Other Pathological Features:

Discussion

Declarations

References

Additional Declarations

Status:

Journal Publication

Version 1