Predictive value of dynamic contrast-enhanced-MRI-based machine learning model for lymphovascular invasion status in node-negative invasive breast cancer

doi:10.21203/rs.3.rs-2951522/v1

Purpose

To retrospectively evaluate breast magnetic resonance imaging (MRI)-based machine learning (ML) model that can preoperatively predict lymphovascular invasion (LVI) status in invasive breast cancer patients with negative axillary lymph nodes (LNs).

Methods

We retrospectively enrolled 280 patients who underwent pre-operative breast MRI and were confirmed with invasive breast cancer by surgical pathology, with absence of axillary LN metastasis. This cohort included 148 LVI-positive lesions and 141 LVI-negative lesions, randomly divided (7:3) into training and validation cohort. The tumor segmentation from the first postcontrast dynamic contrast-enhanced (DCE)-MRI was semi-automatically implemented using 3D Slicer 5.1.0 software, and radiomics features of each lesion were extracted. Least absolute shrinkage and selection operator (LASSO) regression method was used to select the significant radiomics features, and five different ML algorithms were used to build models. A clinical model was also built and integrated with the radiomics models. The predictive performance of models were evaluated by discrimination and compared using DeLong test.

Results

A total of 4 clinico‑radiological parameters and 10 selected radiomics features were selected for model construction. The RF model that was built based on the radiomics and clinico‑radiological feature information exhibited the best predictive performance, yielding high AUC (0.97 for train and 0.82 for validation, respectively). The integrated model significantly outperformed the clinical model, whereas showed no significant difference from the radiomics model.

Conclusions

The RF model integrating radiomics features with clinical information facilitate LVI detection in invasive breast cancer patient with negative axillary LNs preoperatively, which was valuable for clinical decision-making.

Breast cancer

Lymphovascular invasion

Magnetic resonance imaging

Machine learning

Lymphovascular invasion (LVI), infiltration of cancer cells in the lymphatic or vascular intraluminal area at the periphery of the tumor primary site[1], has been broadly accepted as an unfavorable prognostic factor for lymphnode (LN)-negative breast cancer patients[2]. Obviously, LVI is one of the essential steps of lymph node metastasis, therefore LVI-positive tumors have greater potentiality of local recurrence and distant metastasis than LVI-negative ones. It has been proved that the presence of LVI is a crucial indicator of adverse outcome[3; 4]. Accordingly, for those with LVI among LN-negative breast cancer patients, more aggressive treatment decision such as sentinel lymph node biopsy (SLNB) and axillary lymph node dissections is necessarily made in actual clinical practice[5]. However, the status of LVI is currently determined by the surgical pathology specimens, which may bring consequent troubles including delayed treatments, additional surgeries, unnecessary anxiety and poor quality-of-life. Apparently, identifying the presence of LVI as early as possible during the entire treatment is of vital significance. At present, preoperative needle biopsy is ideal choice for most patients to determine the characteristics of the tumors, whereas it is difficult to accomplish the accurate detection of LVI at the partial sampling from core biopsy considering the tissue shrinkage and mechanically-induced cellular displacement[6].

It has been proven that multiparametric MRI is an extraordinarily reliable, comprehensive and accurate diagnostic technique in preoperative staging with the highest sensitivity and fairly great specificity for breast malignancy detection among current imaging patterns[7; 8]. Several studies on breast MRI have indicated that some certain MRI features, such as mass margins, adjacent vessel sign (AVS), peritumoral edema, tumor size, background parenchymal enhancement (BPE), apparent diffusion coefficient (ADC) patterns and contrast enhancement patterns are associated with LVI[2; 7; 9; 10]. Nevertheless, their results are more or less inconsistent due to some subjective factors during the process of features assessment.

Radiomics is an emerging branch of science in recent years. It refers to the high-throughput extraction of a large amount of noninvasive information from images (CT, MRI, PET, etc.) to achieve tumor segmentation, feature extraction and model building, and to assist radiologists to make more accurate diagnosis by means of deeper mining, prediction and analysis of massive data[11]. Several studies have used MRI-based radiomics method for LVI prediction in patients with breast cancer[12–15]. However, those study cohort included LN-positive patients for whom the clinical significance of LVI prediction was not quite large. In addition, to the best of our knowledge, only two study has incorporated the clinico‑radiological features to radiomics signatures for LVI-predictive model establishment[13; 15]. In this study, we assessed the relationship between LVI and integrated variables, including both radiomics and clinico‑radiological features, extracted from a relatively larger sample size than in previous studies and using five different ML methods for model construction to make the results more reliable and accurate.

Therefore, the current study aimed to explore whether a ML model, incorporating MRI-based radiomics features, clinicopathologic characteristics and conventional MRI manifestations, could be used to as a preoperative tool for LVI prediction in invasive breast cancer patient with negative axillary lymph nodes.

Patients

This retrospective study was approved by the institutional review board of our hospital and the informed consent requirement was waived. A total of 1275 consecutive female patients who were diagnosed with breast invasive carcinoma from January 2020 to June 2022, were recruited. The exclusion criteria were detailed in Appendix E1. Finally, a total of 280 patients with 289 lesions were included and were randomly (7:3) grouped into training dataset (n = 202) and validation dataset (n = 87). The overall study design is depicted in Fig. 1.

Image acquisition protocol

MRI examinations were performed in the prone position using a 1.5T scanner (Signa HDxt, GE Healthcare, USA) and 3.0T scanner (Signa Pioneer and Discovery 750, GE Healthcare, USA) with a dedicated four or eight-channel phased-array breast coil. The routine protocols of our center consisted of the following sequences: axial fast spin echo (FSE) T1-weighted imaging (T1WI), axial fat-suppressed spin-echo T2-weighted imaging (T2WI), axial echo-planar diffusion-weighted imaging (DWI), and a sagittal DCE sequence. A total of six dynamic sagittal sequences were acquired once before and five times after the injection of contrast medium (Gd-DTPA, 0.2 mL/kg body weight, flow rate 2.0 mL/s) with 60s intervals, immediately followed by an axial delayed-enhancement sequence. The relevant parameters for each sequence were detailed in Appendix E2.

Image interpretation

All MRI manifestations were analyzed independently by two radiologists with 6 and 8 years of experience in breast imaging diagnosis according to the 2013 MRI Breast Imaging Reporting and Data System (BI-RADS®) lexicon of the American College of Radiology. Both observers were blinded to all histopathological information. The following imaging conventional findings of lesions were evaluated: the type of lesion, tumor size, the location of lesion, tumor margin, time-intensity curve (TIC) patterns, BPE, AVS, and tumor ADC value. All images were sent to the Advantage Workstation (AW 4.6 and AW4.7, GE Healthcare) for further post-processing, including TIC and ADC value measurement, using Functool and READY View software. Circular ROI accounting for approximately 75% of each lesion and avoiding the hemorrhage or necrosis area was drawn and used. The TIC patterns were assessed as the following three types (I II and III): progressive enhancement pattern, plateau pattern, and washout pattern. BPE was divided into two categories: extremely minimal or mild enhancement and moderate or marked enhancement. Peritumoral edema was defined as high signal intensity surrounding the tumor on T2WI[16]. The AVS positive was defined as either one or more vessels entering the lesion can be observed on the T1WI post-enhancement sequences[17]. The largest diameter of tumor on the image section showing the largest lesion was measured and recorded. For the semiquantitative kinetic curve parameters, automatedly computed by Functool software, we measured SI₀ (signal intensity values in the precontrast images), SI₁ (signal intensity values in the first postcontrast images) and SI_max (the max signal intensity values in the postcontrast images). The early enhancement ratio (EER) was calculated as follow: EER= (SI₁–SI₀)/ SI₀×100%. The peak enhancement ratio (PER) was calculated as follow: PER=(SI_max–SI₀)/ SI₀×100%. The time to peak enhancement (TTP) was also obtained. The reliability of the observations was assessed using the intraclass correlation coefficient (ICC). Features with ICCs larger than 0.75 suggested satisfactory reproducibility and were reserved for further analysis.

Histopathological assessment

The expression levels of estrogen receptor (ER), progesterone receptor (PR), human epidermal growth factor receptor 2 (HER2), Ki-67 and presence of IMPCs were pathologically evaluated by according to hematoxylin and eosin (H&E) and immunohistochemical analysis. An expression level > 1% of ER and PR was considered as positive status. An HER2 value of 0 or 1 was considered as negative status and a value of 3 or greater was considered as positive status. If the HER2 value was 2, then fluorescence in situ hybridization (FISH) was required for the further confirmation of HER2 status. The high expression of Ki-67 was ≥ 14% and the low expression of Ki-67 was <14%. All the cases were classified into four immunohistochemical subtypes: luminal A, luminal B, HER2-positive, and triple negative based on 2013 St. Gallen Consensus Conference[18].

Volume of Interest (VOI) segmentation and radiomic feature extraction

The segmentations for all MRI images were performed by one radiologist with 6 years of experience in MRI interpretation, and then 100 randomly selected cases were assigned to another radiologist with 8 years of experience in MRI interpretation. Both radiologists were blinded to the patients’ clinical data. An example of segmentation process is shown in Figure S1.

Feature extraction was performed using the radiomics module under 3D Slicer software. A total of 851 radiomic features were extracted from each ROI. The extracted radiomic features included 107 original features and 744 wavelet-based features (comprising 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), Neighboring gray tone difference matrix (NGTDM), shape-based features).

Radiomics and clinico-radiological feature selection

Feature selection and ML model development were performed in the training cohort, whereas the validation cohort was kept completely independent and invisible until the evaluation of the final model performance. The method of feature standardization (standard deviation) was carried out before feature selection. Two steps were included for selecting robust radiomics features. First, to reduce the risk of overfitting, features with highly pairwise correlations at the level of |r|≥ 0.9 were removed by Pearson’s correlation analysis. Second, we performed univariate selection using SelectKBest method for every feature, and features with p values of < 0.05 were retained for further analysis. Third, the least absolute shrinkage and selection operator (LASSO) was adopted to identify the optimal log (λ) value by 5-fold cross-validation and obtain the robust radiomic features with a nonzero coefficient for the differentiation of LVI and non-LVI group.

The independent clinico-radiological features were selected among the variables found to be significant in univariate analysis by multivariate logistic regression analysis.

Model construction and validation

Five different ML classification algorithms namely, random forest [RF], logistic regression [LR], support vector machine [SVM], eXtreme Gradient Boosting [XGBoost], and stochastic gradient descent [SGD], were used to build model for predicting LVI in the training cohort. The relevant parameters for each classifier were detailed in Appendix E3. The predictive models regarding to radiomics signature (model 1), selected clinico-radiological variables (model 2) and integrated features (combining the above two parameters, model 3) were constructed in the training cohort, and validated in the testing dataset.

The performance of the models was assessed in terms of discrimination, including the receiver operating characteristic (ROC) curve, the area under the ROC curve (AUC), accuracy, sensitivity, specificity, positive prediction value (PPV) and negative prediction value (NPV). The AUCs for the ML models were compared using DeLong test.

Statistical analysis

Statistical analyses were performed with the MedCalc® version 20.0.3 (edCalc Software Ltd, Ostend, Belgium) and R software version 4.2.2 (http://www.Rproject.org). Continuous variables are expressed as the mean ± standard deviation (SD), and categorical variables as frequencies or percentages, unless otherwise specified. The Kolmogorov–Smirnov test was used to test whether data was normally distributed. The continuous variables were compared using the Mann–Whitney U test, whereas dichotomous qualitative variables were assessed using the chi-square test or Fisher’s exact test. Univariate and multivariate logistic regression method were used to determine the association of clinico-pathological and radiological features with LVI. Radiomics feature selection and all model development were carried out on Shukun Medical Research Platform (https://medresearch.shukun.net/project). All tests were performed 2-sided, and P < 0 .05 was considered statistically significant.

Patient and MRI characteristics

The ICCs for all MRI characteristics were greater than 0.75 between the two radiologists. Comparison of clinico-radiological features between LVI-positive and LVI-negative patient group was displayed in Table 1. A total of 280 consecutive patients were included, (median age, 47 years; range, 22–74 years). LVI was found in 140 patients with 148 lesions, and was negative in 140 patients with 141 lesions. In the univariate analysis, patient age, peritumoral edema, TTP, and IMPCs were significantly correlated to LVI (p < 0.05). In the multivariate logistic regression analysis, patient age (odds ratio [OR] = 1.03; 95%confidence interval [CI]: 1.00-1.05; p = 0.02), peritumoral edema (OR = 2.44; 95%CI: 1.18–5.08; p = 0.02), TTP (OR = 1.00; 95%CI: 0.99-1.00; p = 0.04) and IMPCs (OR = 0.35; 95%CI: 0.14–0.87; p = 0.02) remained independent predictors of LVI (Table 2 and Fig. 2).

Table 1

Clinico-pathological and MRI characteristics of breast cancer patients with or without LVI
Characteristic	Overall	LVI (n = 148)	Non-LVI (n = 141)	P value
Age, y	47.00 (40.00–55.00)	46.00 (39.00–53.00)	47.00 (42.00–58.00)	0.034
Menopausal status				0.098
Premenopausal	186 (64.4%)	102 (68.9%)	84 (59.6%)
Postmenopausal	103 (35.6%)	46 (31.1%)	57 (40.4%)
Maximum diameter, cm	2.20 (1.70–3.10)	2.20 (1.70–3.20)	2.30 (1.70–2.80)	0.920
Lesion type				0.417
Mass	230 (79.6%)	115 (77.7%)	115 (81.6%)
NME	59 (20.4%)	33 (22.3%)	26 (18.4%)
TIC pattern				0.499
Type I	2 (0.7%)	2 (1.4%)	0 (0.0%)
Type III& Type II	287 (99.3%)	146 (98.6%)	141 (100.0%)
ADC value (mm²/s)	0.86 (0.79–0.97)	0.87 (0.79–0.95)	0.86 (0.78–0.99)	0.993
Peritumoral edema				0.041
Present	55 (19.0%)	35 (23.6%)	20 (14.2%)
Absent	234 (81.0%)	113 (76.4%)	121 (85.8%)
AVS				0.618
Present	233 (80.6%)	121 (81.8%)	112 (79.4%)
Absent	56 (19.4%)	37 (18.2%)	29 (20.6%)
Intratumoral T2 high signal				0.078
Present	48 (16.6%)	19 (12.8%)	29 (20.6%)
Absent	241 (83.4%)	129 (87.2%)	112 (79.4%)
EER(%)	209.00 (178.75–237.00)	211.00 (177.00-237.00)	205.00 (179.75-236.25)	0.956
PER(%)	213.00 (182.00-239.00)	214.00 (181.00-238.50)	212.00 (184.75–242.00)	0.952
TTP(s)	162.00 (142.00-230.00)	175.00 (148.00-236.00)	154.00 (139.00-214.00)	0.002
BPE				0.076
Minimal or mild	185 (64.0%)	102 (68.9%)	83 (58.9%)
Moderate or marked	104 (36.0%)	46 (31.1%)	58 (41.1%)
DWI rim sign				0.188
Present	159 (55.0%)	87 (58.8%)	69 (48.9%)
Absent	130 (45.0%)	61 (41.2%)	72 (51.1%)
Histology				0.563
IDC	255 (88.2%)	129 (87.2%)	126 (89.4%)
Mixed or other	34 (11.8%)	19 (12.8%)	15 (10.6%)
ER status				0.053
Negative	36 (12.5%)	13 (8.8%)	23 (16.3%)
Positive	253 (87.5%)	135 (91.2%)	118 (83.7%)
PR status				0.479
Negative	73 (25.3%)	40 (27.0%)	33 (23.4%)
Positive	216 (74.7%)	108 (73.0%)	108 (76.6%)
HER2 status				0.982
Negative	242 (83.7%)	124 (83.8%)	118 (83.7%)
Positive	47 (16.3%)	24 (16.2%)	23 (16.3%)
Ki-67 status				0.181
High (≥ 14%)	194 (67.1%)	94 (63.5%)	100 (70.9%)
Low (< 14%)	95 (32.9%)	54 (36.5%)	41 (29.1%)
AR status				0.351
Negative	47 (16.3%)	27 (18.2%)	20 (14.2%)
Positive	242 (83.7%)	121 (81.8%)	121 (85.8%)
Histological grade				0.850
I	14 (4.8%)	4 (2.7%)	10 (7.1%)
II	204 (70.6%)	110 (74.3%)	94 (66.7%)
III	71 (24.6%)	34 (23.0%)	37 (26.2%)
Molecular subtype				0.153
Luminal A	68 (23.5%)	35 (23.6%)	33 (23.4%)
Luminal B	185 (64.0%)	100 (67.6%)	85 (60.3%)
HER2 positive	13 (4.5%)	3 (2.0%)	10 (7.1%)
Triple negative	23 (8.0%)	10 (6.8%)	13 (9.2%)
IMPCs				0.020
Present	26 (9.0%)	19 (12.8%)	7 (5.0%)
Absent	263 (91.0%)	129 (87.2%)	134 (95.0%)
Abbreviations: LVI, lymphovascular invasion; NME, non-mass enhancement; TIC, time-intensity curve; ADC, apparent diffusion coefficient; AVS, adjacent vessel sign; EER, early enhancement ratio; PER, peak enhancement ratio; TTP, time to peak; BPE, background parenchymal enhancement; IDC, invasive ductal carcinoma; ER, estrogen receptor; PR, progesterone receptor; HER-2, human epidermal growth factor receptor 2; AR, androgen receptor; IMPCs, invasive micropapillary components

Table 2

Univariable and multivariable logistic regression analyses for clinico-pathological and MRI characteristics in primary cohort (n = 289)
Variables	Univariable analysis		Multivariable analysis
Variables	OR (95% CI)	P value	OR (95% CI)	P value
Age, y	0.974 (0.951, 0.997)	0.024	0.958 (0.934, 0.983)	0.001
Peritumoral edema	2.117 (1.063, 4.213)	0.033	2.474 (1.191, 5.142)	0.015
TTP(s)	1.004 (1.000, 1.007)	0.037	1.004 (1.001, 1.007)	0.024
IMPCs	2.820 (1.147, 6.933)	0.024	3.205 (1.235, 8.317)	0.017
Abbreviations: TTP, time to peak; IMPCs, invasive micropapillary components

Radiomics feature selection

In this study, a total of 851 radiomics features for each lesion were extracted, which included 107 original features and 744 wavelet-based features. A total of 282 features remained after feature preselection with pearson correlation analysis. Then, with single variable feature selection method (SelectKBest), 117 features remained. After further feature dimensionality reduction by LASSO regression algorithm, 10 significant radiomics features were finally selected for the subsequent establishment of ML models (Figure S2 and Fig. 3).

Predictive models assessment

The radiomics predictive models were accordingly built using different ML algorithms, which showed good diagnostic performance in the training and testing sets and the corresponding ROC, AUC, accuracy, sensitivity, specificity, PPV and NPV were displayed in Table 3 and Fig. 4. The best predictive performance validated in the testing dataset was observed in RF model (AUC = 0.816; 95%CI, 0.719–0.891).

Table 3

Diagnostic performance of radiomics models based on 5 different ML algorithms
Models	AUC	95%CI		Sensitivity	Specificity	Accuracy	PPV	NPV
Models	AUC	Lower	Upper	Sensitivity	Specificity	Accuracy	PPV	NPV
LR
Training set	0.908	0.868	0.949	0.737	0.971	0.856	0.961	0.794
Validation set	0.810	0.714	0.906	0.691	0.867	0.782	0.829	0.750
RF
Training set	0.973	0.956	0.990	0.879	0.942	0.911	0.936	0.890
Validation set	0.816	0.720	0.912	0.714	0.800	0.759	0.769	0.750
SGD
Training set	0.909	0.869	0.949	0.717	0.981	0.852	0.973	0.783
Validation set	0.813	0.719	0.906	0.619	0.867	0.747	0.813	0.709
SVM
Training set	0.908	0.867	0.949	0.727	0.981	0.856	0.973	0.789
Validation set	0.802	0.705	0.899	0.595	0.889	0.747	0.833	0.702
XGBoost
Training set	1.000	1.000	1.000	1.000	1.000	1.000	1	1
Validation set	0.804	0.702	0.905	0.714	0.822	0.770	0.790	0.755
Abbreviations: AUC, area under the ROC curves; CI, confidence interval; PPV, positive prediction value; NPV, negative prediction value; LR, logistic regression; RF, random forest; SGD, stochastic gradient descent; SVM, support vector machine; XGBoost, Xtreme Gradient Boosting

The models based on selected clinico-radiological variables applied to five ML algorithms had relatively good performance in predicting LVI, and their ROC, AUC, accuracy, sensitivity, specificity, PPV and NPV were shown in Table S1. The best predictive performance validated in the testing dataset was also observed in RF model (AUC = 0.720; 95%CI, 0.611–0.830).

In the ML models based on combined features, including selected 10 radiomics signatures and 4 clinico-radiological variables, the SGD model performed poorer than both radiomics model (AUC: 0.500 vs. 0.813, 95%CI, 0.219–0.406, Z = 6.542, p <0.0001) and clinical model (AUC: 0.500 vs. 0.627, 95%CI, 0.007–0.246, Z = 2.077, P = 0.038). The other conjoint models outperformed those based on clinico-radiological variables alone to predict LVI. The best integrated model was RF model (AUC = 0.833; 95%CI, 0.742–0.925, Table 4), and the the variables (selected clinico-radiological variables and the Radscore calculated by radiomics RF model) are ranked in importance order according to the RF model (Fig. 5). Then, the performance of model 1,2 and 3 in the optimal ML classifier (RF algorithm) is displayed in Table 5 and Fig. 6. In the testing cohort, the Delong test showed statistical difference in AUCs between model 2 and model 3 (p = 0.047) but no statistical difference between model 1 and model 3 (p = 0.679).

Table 4

Diagnostic performance of combined models based on radiomics and clinico-radiological variables applied to 5 different ML algorithms
Models	AUC	95%CI		Sensitivity	Specificity	Accuracy	PPV	NPV
Models	AUC	Lower	Upper	Sensitivity	Specificity	Accuracy	PPV	NPV
LR
Training set	0.909	0.869	0.950	0.798	0.922	0.861	0.908	0.826
Validation set	0.805	0.708	0.901	0.7381	0.800	0.770	0.775	0.766
RF
Training set	0.972	0.954	0.990	0.919	0.903	0.911	0.901	0.921
Validation set	0.833	0.742	0.925	0.810	0.733	0.770	0.739	0.805
SGD
Training set	0.500	0.500	0.500	1.000	0.000	0.490	0.490	0.000
Validation set	0.500	0.500	0.500	1.000	0.000	0.483	0.483	0.000
SVM
Training set	0.889	0.843	0.934	0.727	0.932	0.832	0.911	0.781
Validation set	0.797	0.699	0.895	0.667	0.822	0.747	0.778	0.726
XGBoost
Training set	0.995	0.990	1.000	0.990	0.952	0.970	0.952	0.990
Validation set	0.821	0.727	0.915	0.738	0.689	0.713	0.689	0.738
Abbreviations: AUC, area under the ROC curves; CI, confidence interval; PPV, positive prediction value; NPV, negative prediction value; LR, logistic regression; RF, random forest; SGD, stochastic gradient descent; SVM, support vector machine; XGBoost, Xtreme Gradient Boosting

Table 5

Comparation on the performance of RF models (model 1, model 2 and model 3) in identifying LVI in the training and validation cohorts
Models	AUC	95%CI		Sensitivity	Specificity	Accuracy	P value
Models	AUC	Lower	Upper	Sensitivity	Specificity	Accuracy	P value
Model 1							0.679*
Training set	0.973	0.956	0.990	0.879	0.942	0.911
Validation set	0.816	0.720	0.912	0.714	0.800	0.759
Model 2							0.047*
Training set	0.845	0.793	0.897	0.778	0.767	0.772
Validation set	0.720	0.611	0.830	0.762	0.667	0.713
Model 3
Training set	0.972	0.954	0.990	0.919	0.903	0.911
Validation set	0.833	0.742	0.925	0.810	0.733	0.770
Abbreviations: AUC, area under the ROC curves; CI, confidence interval
*DeLong test for comparing with model 3 in each validation set

In this study, we investigated the relationship between LVI and invasive breast cancer in node-negative patients and constructed MRI-based radiomics models, which had a highly robust ability to predict LVI. Five different ML algorithms were used for model construction based on the 10 selected radiomics signatures, and the RF model outperformed the other models on the power of identifying LVI status. Moreover, the combined model integrating clinico-radiological characteristics with radiomics signatures revealed a greatly better predictive performance than the clinical model. It was expected to offer a relatively accurate and objective approach for preoperative LVI prediction.

LVI has been proven as an important pathological predictor of the metastasis potential and a strong prognostic factor for invasive breast cancer[4; 16; 19; 20]. However, it is always difficult to confirm the presence of LVI in preoperative biopsy material; therefore, to predict the LVI status from preoperative imaging data is expected to provide huge clinical value. Some studies had showed LVI was associated with age, tumor size, mass margin, positive lymph nodes, BPE, TIC patterns, ADC value, peritumoral edema, AVS, DWI rim sign and histologic grade[1; 2; 7; 10; 21]. Our study found the risk for LVI increased with lower mean age. The mean patient age with LVI was smaller than that in patients without LVI. A previous study revealed that age was a high-risk factor for LVI in invasive breast cancer[22]. Yang et al. [23]stated that premenopausal status was associated with an increased risk (4.59-fold) of early-onset breast cancer increased the risk 4.59-fold in early-onset BC (< 40 years). Among premenopausal women, early-onset breast cancer were more aggressive than those among older women because blood estrogen and progesterone levels gradually decrease with age[24].

We also identified peritumoral edema as a predictor of LVI; those patients without peritumoral edema had higher risk for LVI, which was consistent with the previous study results[15; 21]. Peritumoral edema was found in a significantly higher proportion in patients with more biologically aggressive tumors[16]. The biological mechanism underlying the association between peritumoral edema and LVI remains unknown. The possible reason for peritumoral edema may be the increase of vascular permeability and obstruction of lymphatic drainage by tumor emboli[25]. Additionally, TTP, a semiquantitative DCE-MRI parameter, was found to be an independent predictive factor of LVI. We revealed that patients with longer TTP had a higher probability of positive LVI, which might be related with the focal reduced blood flow cause by tumor emboli. Numerous studies have shown that TTP is a diagnostic indicator of breast malignant lesions[26; 27] and an independent predictive factor of pathologic complete response (pCR) after neoadjuvant chemotherapy (NAC)[28; 29]. One recent study analysed the relationship between TTP and LVI, but found no significant differences between patients with and without LVI[30]. Especially, our study also identified IMPCs as a predictive indicator for LVI. To the best of our knowledge, this is the first study to include IMPCs to construct a predictive model for detecting LVI status in invasive breast cancer patients. IMPC is a rare subtype of epithelial tumor of the breast listed in the 2003 World Health Organization (WHO) histologic classification of tumors of the breast[31]. Previous studies demonstrated that breast cancers with IMPCs had a relatively increased incidence of LVI[32]. LVI has also been found to be more common among patients with IMPC (pure or mixed with IDC), as reported by Tang et al[33], with 14.7% versus only 0.1% in the IDC group, and 94.7% versus 71.9% in the control group being reported by Gokce et al[34]. In our study, we found higher frequent IMPC in patients with LVI than without LVI, which was consistent with the previous study[35].

A number of studies confirmed that DCE-MRI-based radiomics ML model had a robust power for predicting LVI status of invasive breast cancer patients[12–15]. In their studies, the cohorts included both node-positive and node-negative patients. While in the current study, only node-negative patients were recruited. Since those with axillary LN metastasis, considered as an unfavorable prognostic factor, are usually recommended for axillary node dissection in clinical practice[36]. Therefore, the detection of LVI mildly affects the intended clinical management. Zhang et al[15] identified that the proposed nomogram, incorporating MRI-based radiomics signature and MRI-reported peritumoral edema, achieved a satisfactory preoperative prediction of LVI and clinical outcomes in IDC patients. One recent study used SVM classifier to establish a prediction model of LVI based on ADC radiomic signature, and its AUC was 0.77 in the test set[14]. Liu et al[13] confirmed that the DCE-MRI based radiomics model using multivariate logistic regression method could significantly improve the performance for discriminating LVI-positive from LVI-negative lesions. Another study found that ML-based radiomics model based on 3D segmentation of ADC maps could be used to predict LVI status in breast cancer patients[12]. In their study, the best model performance was achieved in the tuning and boosting RF model, with AUC of 0.726 and 0.732 in the training and test datasets, respectively. In our study, we used a larger sample size and compared five different ML algorithms (including LR, RF, SVM, SGD and XGBoost) to gain relatively more reliable and objective results. For the radiomics signature, one original signature and nine wavelet-based features were selected. Wavelet features can efficiently provide multi-frequency information of the tumor in time-frequency domain[37]. In our study, all these wavelet-based features were significantly higher in the LVI-positive cohort than those in the LVI-negative cohort, which was consistent with the previous study[15].

However, Several limitations of this study should be pointed out. Firstly, this was a retrospective single-center study, indicating the necessity to a large-scale prospective multicenter study to further validate the effectiveness of the proposed radiomics model. Secondly, only the first phase after enhancement were used to extract radiomics features in this study, calling for other sequences such as T2W, DWI images and ADC maps to be used for features extraction in further studies. Thirdly, since our institution originally designed a sagittal DCE sequence for breast MRI examination, the functional tumor volume (FTV) of the lesions could not be obtained. Similarly, our study only analyzed the semiquantitative kinetic curve parameters for LVI prediction. The quantitative pharmacokinetic parameters (such as ktrance, kep, and Ve) were failure to gained with our equipments. Further exploration for these parameters are needed if conditions permit.

In conclusion, radiomics features derived from DCE-MRI are robust biomarkers for predicting LVI. The combined ML model, incorporating radiomics signature and clinico-radiological-based variables, exhibited a highly acceptable predictive efficacy for LVI status in node-negative invasive breast cancer patients.

ADC Apparent diffusion coefficient

AR Androgen receptor

AUC Area under the curve

AVS Adjacent vessel sign

BI-RADS® Breast Imaging Reporting and Data System

BPE Background parenchymal enhancement

EER Early enhancement ratio

ER Estrogen receptor

HER2 Human epidermal growth factor receptor 2

ICC Interclass correlation coefficient

IDC Invasive ductal carcinoma

IMPCs Invasive micropapillary components

LASSO Least absolute shrinkage and selection operator

LVI Lymphovascular invasion

ML Machine learning

NME Nonmass enhancement

PER Peak enhancement ratio

PR Progesterone receptor

ROC Receiver operating characteristic

SLNB Sentinel lymph node biopsy

TIC Time-intensity curve

TTP Time to peak

Funding

This work was supported by the National Natural Science Foundation of China [grant numbers 82172025]; Chinese National Key Research and Development Project [grant number 2021YFC2500400]

Conflict of Interest

The authors of this manuscript declare no relationships with any companies, whose products or services may be related to the subject matter of the article.

Igarashi T, Furube H, Ashida H, Ojiri H (2018) Breast MRI for prediction of lymphovascular invasion in breast cancer patients with clinically negative axillary lymph nodes. Eur J Radiol 107:111-118
Choi BB (2021) Dynamic contrast enhanced-MRI and diffusion-weighted image as predictors of lymphovascular invasion in node-negative invasive breast cancer. World J Surg Oncol 19:76
Schoppmann SF, Bayer G, Aumayr K et al (2004) Prognostic value of lymphangiogenesis and lymphovascular invasion in invasive breast cancer. Ann Surg 240:306-312
Zhang S, Zhang D, Gong M, Wen L, Liao C, Zou L (2017) High lymphatic vessel density and presence of lymphovascular invasion both predict poor prognosis in breast cancer. BMC Cancer 17:335
Ryu JM, Lee SK, Kim JY et al (2017) Predictive Factors for Nonsentinel Lymph Node Metastasis in Patients With Positive Sentinel Lymph Nodes After Neoadjuvant Chemotherapy: Nomogram for Predicting Nonsentinel Lymph Node Metastasis. Clin Breast Cancer 17:550-558
Cheung SM, Husain E, Mallikourti V, Masannat Y, Heys S, He J (2021) Intra-tumoural lipid composition and lymphovascular invasion in breast cancer via non-invasive magnetic resonance spectroscopy. Eur Radiol 31:3703-3711
Li J, Mo Y, He B et al (2019) Association between MRI background parenchymal enhancement and lymphovascular invasion and estrogen receptor status in invasive breast cancer. Br J Radiol 92:20190417
Mann RM, Cho N, Moy L (2019) Breast MRI: State of the Art. Radiology 292:520-536
Okuma H, Sudah M, Kettunen T et al (2020) Peritumor to tumor apparent diffusion coefficient ratio is associated with biologically more aggressive breast cancer features and correlates with the prognostication tools. PLoS One 15:e0235278
Ouyang FS, Guo BL, Huang XY et al (2019) A nomogram for individual prediction of vascular invasion in primary breast cancer. Eur J Radiol 110:30-38
Gillies RJ, Kinahan PE, Hricak H (2016) Radiomics: Images Are More than Pictures, They Are Data. Radiology 278:563-577
Kayadibi Y, Kocak B, Ucar N, Akan YN, Yildirim E, Bektas S (2022) MRI Radiomics of Breast Cancer: Machine Learning-Based Prediction of Lymphovascular Invasion Status. Acad Radiol 29 Suppl 1:S126-S134
Liu Z, Feng B, Li C et al (2019) Preoperative prediction of lymphovascular invasion in invasive breast cancer with dynamic contrast-enhanced-MRI-based radiomics. J Magn Reson Imaging 50:847-857
Nijiati M, Aihaiti D, Huojia A et al (2022) MRI-Based Radiomics for Preoperative Prediction of Lymphovascular Invasion in Patients With Invasive Breast Cancer. Front Oncol 12:876624
Zhang J, Wang G, Ren J et al (2022) Multiparametric MRI-based radiomics nomogram for preoperative prediction of lymphovascular invasion and clinical outcomes in patients with breast invasive ductal carcinoma. Eur Radiol 32:4079-4089
Cheon H, Kim HJ, Kim TH et al (2018) Invasive Breast Cancer: Prognostic Value of Peritumoral Edema Identified at Preoperative MR Imaging. Radiology 287:68-75
Dietzel M, Baltzer PA, Vag T et al (2010) The adjacent vessel sign on breast MRI: new data and a subgroup analysis for 1,084 histologically verified cases. Korean J Radiol 11:178-186
Untch M, Gerber B, Harbeck N et al (2013) 13th st. Gallen international breast cancer conference 2013: primary therapy of early breast cancer evidence, controversies, consensus - opinion of a german team of experts (zurich 2013). Breast Care (Basel) 8:221-229
Ma Q, Dieterich LC, Detmar M (2018) Multiple roles of lymphatic vessels in tumor progression. Curr Opin Immunol 53:7-12
Rakha EA, Martin S, Lee AH et al (2012) The prognostic significance of lymphovascular invasion in invasive breast carcinoma. Cancer 118:3670-3680
Cheon H, Kim HJ, Lee SM et al (2017) Preoperative MRI features associated with lymphovascular invasion in node-negative invasive breast cancer: A propensity-matched analysis. J Magn Reson Imaging 46:1037-1044
Zhang C, Liang Z, Feng Y, Xiong Y, Manwa C, Zhou Q (2022) Risk Factors for Lymphovascular Invasion in Invasive Ductal Carcinoma Based on Clinical and Preoperative Breast MRI Features: a Retrospective Study. Acad Radiol. 10.1016/j.acra.2022.10.029
Yang PJ, Hou MF, Ou-Yang F, Tsai EM, Wang TN (2022) Association of early-onset breast cancer with body mass index, menarche, and menopause in Taiwan. BMC Cancer 22:259
King V, Gu Y, Kaplan JB, Brooks JD, Pike MC, Morris EA (2012) Impact of menopausal status on background parenchymal enhancement and fibroglandular tissue on breast MRI. Eur Radiol 22:2641-2647
Uematsu T (2015) Focal breast edema associated with malignancy on T2-weighted images of breast MRI: peritumoral edema, prepectoral edema, and subcutaneous edema. Breast Cancer 22:66-70
Fan WX, Chen XF, Cheng FY et al (2018) Retrospective analysis of the utility of multiparametric MRI for differentiating between benign and malignant breast lesions in women in China. Medicine (Baltimore) 97:e9666
Liu F, Kornecki A, Shmuilovich O, Gelman N (2011) Optimization of time-to-peak analysis for differentiating malignant and benign breast lesions with dynamic contrast-enhanced MRI. Acad Radiol 18:694-704
Panzeri MM, Losio C, Della Corte A et al (2018) Prediction of Chemoresistance in Women Undergoing Neo-Adjuvant Chemotherapy for Locally Advanced Breast Cancer: Volumetric Analysis of First-Order Textural Features Extracted from Multiparametric MRI. Contrast Media Mol Imaging 2018:8329041
Li Y, Chen Y, Zhao R et al (2022) Development and validation of a nomogram based on pretreatment dynamic contrast-enhanced MRI for the prediction of pathologic response after neoadjuvant chemotherapy for triple-negative breast cancer. Eur Radiol 32:1676-1687
Lai T, Chen X, Yang Z et al (2022) Quantitative parameters of dynamic contrast-enhanced magnetic resonance imaging to predict lymphovascular invasion and survival outcome in breast cancer. Cancer Imaging 22:61
Walker RA (2010) World Health Organization Classification of Tumours. Pathology and Genetics of Tumours of the Breast and Female Genital Organs. Histopathology 46:229-229
Chen H, Wu K, Wang M, Wang F, Zhang M, Zhang P (2017) Invasive micropapillary carcinoma of the breast has a better long-term survival than invasive ductal carcinoma of the breast in spite of its aggressive clinical presentations: a comparison based on large population database and case-control analysis. Cancer Med 6:2775-2786
Tang SL, Yang JQ, Du ZG et al (2017) Clinicopathologic study of invasive micropapillary carcinoma of the breast. Oncotarget 8:42455-42465
Gokce H, Durak MG, Akin MM et al (2013) Invasive micropapillary carcinoma of the breast: a clinicopathologic study of 103 cases of an unusual and highly aggressive variant of breast carcinoma. Breast J 19:374-381
Kuroda H, Sakamoto G, Ohnisi K, Itoyama S (2004) Clinical and pathologic features of invasive micropapillary carcinoma. Breast Cancer 11:169-174
Marino MA, Avendano D, Zapata P, Riedl CC, Pinker K (2020) Lymph Node Imaging in Patients with Primary Breast Cancer: Concurrent Diagnostic Tools. Oncologist 25:e231-e242
Qin Q, Li J, Zhang L, Yue Y, Liu C (2017) Combining Low-dimensional Wavelet Features and Support Vector Machine for Arrhythmia Beat Classification. Sci Rep 7:6067

SupplementaryInformation.docx

Predictive value of dynamic contrast-enhanced-MRI-based machine learning model for lymphovascular invasion status in node-negative invasive breast cancer

Status:

Version 1

Abstract

Purpose

Methods

Results

Conclusions

Figures

Introduction

Materials and methods

Patients

Image acquisition protocol

Image interpretation

Histopathological assessment

Volume of Interest (VOI) segmentation and radiomic feature extraction

Radiomics and clinico-radiological feature selection

Model construction and validation

Statistical analysis

Results

Patient and MRI characteristics

Radiomics feature selection

Predictive models assessment

Discussion

Abbreviations

Declarations

References

Supplementary Files

Status:

Version 1