Machine Learning with Multiparametric MRI-based Radiomics Models for Preoperative Prediction of Ki-67 Status in Luminal Breast Cancer

doi:10.21203/rs.3.rs-3557175/v1

Download PDF

Research Article

Machine Learning with Multiparametric MRI-based Radiomics Models for Preoperative Prediction of Ki-67 Status in Luminal Breast Cancer

https://doi.org/10.21203/rs.3.rs-3557175/v1

This work is licensed under a CC BY 4.0 License

Version 1

posted

You are reading this latest preprint version

Background

The main objective of the study was to determine whether multiparametric MRI (mpMRI) radiomics models supported by machine learning could preoperatively predict Ki-67 status in luminalbreast carcinoma.

Methods

Between 2018 and 2021, patients with luminal breast cancer who underwent mpMRI in our institution were retrospectively enrolled. The Ki-67 status was analyzed by biopsy preoperatively. Radiomics features were extracted from the T2WI, DCE, DWI, and ADC images, and mpMRI features were derived from four MRI sequences. A prediction model was developed by training the logistic regression (LR), multilayer perceptron (MLP), random forest (RF), and support vector machine (SVM) machine learning classifiersaccording to the radiomic characteristics. A clinical-radiomic nomogram was constructed by integrating mpMRI radiomic features and routine clinical MRI variables, followed by calibration and decision curve analyses.

Results

A total of 140 patients (85 with high and 55 with low Ki-67 expression) were enrolled. Compared to the DCE-, DWI-, and ADC-based radiomic signatures, the T2WI-based radiomic signature exhibited high prediction quality with AUCs of 0.87, 0.92, 0.92, and 0.89 for the four classification algorithms (LG, RF, MLP, SVM), respectively (all p<0.05). The mpMRI radiomic signature also showed high quality with AUCs of 0.92, 0.89, 0.92, and 0.92 for the four algorithms (all p<0.05). A prediction clinical-radiomicnomogram was constructed with training and validation set AUCs of 0.93 (0.90-0.96) and 0.92 (0.89-0.95), respectively.

Conclusion

T2-based and mpMRI-based radiomics models combined with advanced machine learning classifiers could assist in the preoperative individual-specific prediction of Ki-67 status in luminalbreast carcinoma.

Breast cancer

Ki-67 antigen

Magnetic resonance imaging

Machine learning

Radiomic

Breast cancer is one of the primary reasons behind mortality due to cancer in women[1]. It is a heterogeneous disease with several molecular subtypes. The 12th St Gallen International Breast Cancer Conference (2011) defined progesterone receptor (PR), estrogen receptor (ER), human epidermal growth factor receptor 2 (HER2) oncogene, and Ki-67 labeling index as the classification criteria for identifying tumor subtypes[2]. Based on molecular portraits, breast cancer was categorized into four molecular subgroups: luminal A (ER, PR positive, HER2 negative, Ki-67 low), luminal B (HER2 negative) (ER, PR positive, HER2 negative, Ki-67 high), luminal B (HER2 positive) (ER, PR positive, any Ki-67), HER2 overexpressed (HER2 overexpressed, ER and PR absent), and basal-like (ER and PR absent, HER2 negative). Different subtypes exhibit varying prognoses and responses to treatment[3–5]. As a result, breast cancer management has evolved into individualized conversations and treatment plans rather than a "one size fits all" approach[6].

Luminal cancer accounts for 73.2% of all breast cancers[7]. Ki-67 is a proliferating biomarker whose levels are a primary indicator of tumor cell proliferation[8–11]. Luminal breast cancer subtypes A and B (HER2 negative) are distinguished based on a 14% Ki-67 cutoff value[2]. The luminal A type, characterized by milder aggressiveness and better prognosis, is primarily treated with endocrine therapy alone. Conversely, the highly proliferative luminal B (HER2 negative) type, demonstrating a more aggressive nature and a poorer prognosis, necessitates a combined approach involving both endocrine therapy and chemotherapy[1, 2, 12].

Currently, Ki-67 assessment predominantly relies on histopathological examination based on preoperative core needle biopsy, which often fails to comprehensively evaluate tumor heterogeneity [13, 14]. This limitation hampers accurate tumor typing, treatment plan optimization, and prognosis assessment. Radiomics is a high-throughput process involving the extraction and analysis of computationally derived features from MR images, enabling the quantification of differences between tumor tissues and providing an overall assessment of breast tumors[15, 16]. Nomograms are graphical instruments that streamline the prediction-making process and find extensive use in cancer prognosis. It effectively transforms complex predictive models into a singular, easy-to-comprehend visual depiction[17]. Decision curve analysis (DCA) is a method to evaluate the clinical performance of prediction models by calculating the "net benefit" [18].

At present, there are several studies utilizing apparent diffusion coefficient (ADC) values[19], T2-weighted imaging (T2WI) and dynamic contrast-enhanced (DCE) radiomic features[20–22] and multiparametric MRI (mpMRI) combined with machine learning methods[23, 24] to predict the Ki-67 status preoperatively. However, there are few studies that combine mpMRI radiomic features with various machine learning methods to predict Ki-67 status. In the present study, we attempted to analyze whether prediction signatures based on four machine learning methods using clinical routine imaging parameters and mpMRI radiomic characteristics from T2WI, DCE, ADC and diffusion-weighted imaging (DWI) can preoperatively predict Ki-67 expression in luminal breast carcinoma.

The institutional review board approved our study (No. 2018-L-23) while waiving the need for written informed consent.

Patient enrollment

We retrospectively reviewed the health-related information from the database of our institution between January 2018 and October 2021. Finally, 320 women satisfying the following inclusion criteria were chosen: (1) MRI performed within two weeks before biopsy; (2) initial diagnosis of unilateral primary breast cancer based on preoperative percutaneous core needle biopsy; and (3) confirmed luminal cancer based on immunohistochemistry (ER or PR positive and HER2 negative) based on biopsy specimens. Ultimately, we removed 180 patients based on the following criteria: (1) MRI examination not performed within two weeks before biopsy (n = 120); (2) missing values of Ki-67 (n = 22); (3) nonmass or multifocal breast cancer (n = 18); and (4) inadequate breast MR images (lacking T2WI, DWI, ADC, DCE) or unclear image quality affecting lesion segmentation and evaluation (n = 20). Ultimately, 140 women were consecutively enrolled (mean age with a standard deviation of 51 years ± 12; age range: 26–77 years). Based on the St. Gallen consensus[8], we divided the participants into two subgroups: (1) a high Ki-67 expression group (≥ 14%, luminal B (HER2-negative) type, n = 85) and (2) a low Ki-67 expression group (< 14%, luminal A type, n = 55) (Fig. 1).

MRI examination

MRI was performed using a 3.0-T scanner (Achieva; Philips Healthcare, best, Netherland) in the axial orientation with a specific coil of a seven-channel breast array. All the imaging sections were obtained by putting the patients in a prone position. Bilateral axial T2-weighted images with fat suppression were acquired (TR, 5375 ms; TE, 65 ms; FOV, 340 mm × 340 mm; flip angle, 90°, matrix, 620 × 303; slice thickness, 3 mm; space, 0 mm). DWI was acquired utilizing echo-planar imaging in the axial plane with fat suppression (TR/TE, 5417 ms/72 ms; FOV, 320 mm×320 mm; flip angle, 90°, matrix, 96 × 126; slice thickness, 3 mm; space, 0 mm; b values, 0 and 1000 s/mm2). A 3D T1-weighted gradient-echo sequence, which was rapidly spoiled, was used to obtain one precontrast and five postcontrast dynamic series by suppressing the fat (TR/TE, 5 ms/2 ms; FOV 340 mm × 340 mm; matrix, 436 × 436; flip angle, 12°; slice thickness, 1 mm; space, 0 mm) by intravenously injecting 0.1 mmol/kg gadoterate (Gadovist, Bayer Healthcare).

Routine clinical MRI evaluation

Two breast radiologists, with 15 and 5 years of experience, independently and blindly assessed all the MR images. They utilized the 2013 Breast Imaging Reporting and Data System MR lexicon for their assessment. Routine clinical variables included patient age, tumor maximum diameter, shape, invasion (skin, nipple, pectoral muscle, chest wall), T2 signal, ADC value, first-minute enhancement rate, maximum enhancement rate, time to peak (Tmax), and time-signal intensity curves (TICs).

Radiomic feature extraction

T2WI, DCE (the 1st phase), DWI (b = 1000), and ADC datasets were incorporated into the 3D-Slicer V4.11.2 free software (https://WWW.slicer.org/), which is typically used for processing images[25]. The same two radiologists performed a semiautomated delineation of each slice of the lesions (including necrosis and liquefaction, which showed intertumoral hyperintensity on T2WI). The regions of interest (ROIs) were delineated using T2WI, DWI and DCE sequences. Subsequently, the ROIs extracted from the DWI sequence were replicated onto the corresponding ADC maps. Radiomic features were extracted from four ROIs using the PyRadiomics package in Python version 3.7 (https://www.python.org/). We extracted 851 features from each segmented sequence of the tumor 3D-ROI (DCE, DWI, T2WI, and ADC). The features had 162 first-order, 14 shapes, and 675 textures, primarily derived from the maximum short axis, kurtosis, and gray-level matrices of co-occurrence, dependency, run length, etc.

Selection of characteristic features and development of radiomic signatures

To reduce redundant features and avoid overfitting, three procedures were sequentially executed to determine the best features for predicting the Ki-67 status in luminal breast cancer. First, the independent sample t test helped eliminate the radiomic features without statistically significant differences between the two groups. Subsequently, the Z score method was utilized to separately normalize the selected radiomic features from the ROIs of each MRI sequence. Then, the LASSO algorithm was used to extract the features with nonzero coefficients at a specific coefficient lambda. Fivefold cross-validation was used to determine the optimal coefficient lambda in the LASSO method.

Establishment and evaluation of machine learning classifiers

We randomly divided the data into training and validation sets at a 7:3 ratio (70% training and 30% validation). (1) Building and validating the routine clinical MRI signature. We developed signatures using the training set with clinical and routine MRI characteristics, and features were incorporated into further selection at a significance of p < 0.05. Next, four classification algorithms (logistic regression (LR), random forest (RF), support vector machine (SVM), and multilayer perceptron (MLP)) were selected for model training and validation. (2) MRI radiomic signature development and authentication. The training set was initially evaluated for predicting Ki-67 expression and then validated using 5-fold cross-validation in the inner loop. In the training set, we established signatures using radiomic features extracted from the T2WI, DCE, DWI and ADC images, as well as mpMRI radiomic features derived from four MRI sequences. Next, the four classification algorithms (LR, RF, MLP, SVM) were used for model training and validation.

Constructing the clinical-radiomic nomogram and decision curve analysis

A predictive clinical-radiomic nomogram model was formulated by constructing multivariate analysis and regression modeling using the Rad-score combined with routine clinical MRI variables. The Rad score was calculated from mpMRI radiomic features and using the intercept and coefficients obtained from LR, as shown in the following formula:

Rad-score = 1.1998 − 0.03480*F1-0.3029*F2 + 0.3678*F3 + 0.2167*F4 + 0.3736*F5 + 0.6594*F6 + 0.5211*F7 + 0.6672*F8 + 0.1984*F9-0.2013*F10 + 0.2827*F11

F1 = wavelet-LLH_gldm_SmallDependenceLowGrayLevelEmphasis (ADC)

F2 = wavelet-HLL_gldm_SmallDependenceLowGrayLevelEmphasis (DCE)

F3 = wavelet-HHH_gldm_DependenceNonUniformity (DCE)

F4 = wavelet-LHL_glcm_Correlationoriginal_shape_MinorAxisLength(T2WI)

F5 = wavelet-LHH_glcm_Imc2original_shape_MinorAxisLength(T2WI)

F6 = wavelet-LHH_gldm_DependenceVarianceoriginal_shape_MinorAxisLength(T2WI)

F7 = wavelet-LHH_glrlm_LongRunLowGrayLevelEmphasisoriginal_shape_MinorAxisLength(T2WI)

F8 = wavelet-HLH_glcm_Correlationoriginal_shape_MinorAxisLength(T2WI)

F9 = wavelet-HLH_glrlm_LongRunEmphasisoriginal_shape_MinorAxisLength(T2WI)

F10 = wavelet-HHH_firstorder_Meanoriginal_shape_MinorAxisLength(T2WI)

F11 = wavelet-HHH_glcm_Idnoriginal_shape_MinorAxisLength(T2WI)

Then, calibration curves were used to evaluate the performance of the nomogram in the training and validation sets. The clinical usefulness of the nomogram was evaluated using clinical decision curve analysis (DCA). Three decision curves were plotted based on the routine clinical variables, radiomic signature and nomogram, and the net benefits of each decision curve were calculated. Python 3.7 was used for conducting all the analyses. We implemented the features and classifiers using the Scikit-learn 1.0.2 package. Figure 2 presents a concise flowchart outlining the workflow of this study.

Statistical analysis

We analyzed and compared the categorical groups with the χ² or Fisher’s exact tests. Additionally, we compared two groups of continuous variables with the independent-sample t test. ROC analysis helped predict the signature performance. The area under the ROC curve (AUC) facilitated the sensitivity and specificity evaluation for every signature. We compared the AUCs for two ROC curves using the DeLong test. Multivariate LR models were constructed for the nomogram. Statistical significance was determined for all analyses with a two-sided p value threshold < 0.05. We conducted the statistical analyses with SPSS version 26.0 and R software (version 4.2.1, http://www.r-project.org).

Routine clinical MRI signature

The clinical and routine MRI features are characterized in Table 1. Compared to the low Ki-67 expression group, the groups with elevated Ki-67 levels showed more regular intertumoral hyperintensity or mixed intensity on T2WI, lobulated shape, spicules of margin, adjacent invasion, larger tumor size, and a more significant washout TIC component (p < 0.05). No substantial differences between the two groups could be observed in age, ADC levels, first-minute enhancement rate, maximum enhancement rate, or Tmax (p = 0.351, 0.985, 0.389, 0.650, and 0.150, respectively; Table 1).

Table 1

Routine Clinical MRI Characteristics
Clinical/MRI feature	Low Ki-67 expression group (n = 55)	High Ki-67 expression group (n = 85)	P value
Age^#	52.35 ± 11.851	49.95 ± 11.440	0.351
Morphological feature
Maximum diameter^#	2.164 ± 1.153	2.544 ± 1.277	0.044*
Spiculated margin			0.043*
present	24(43.6)	34(40.0)
absent	31(56.4)	51(60.0)
Lobulated shape			< 0.001*
present	42(76.4)	62(72.9)
absent	13(23.6)	23(27.1)
Invasion of adjacent tissues			< 0.001*
present	12(21.8)	26(30.6)
absent	43(78.2)	59(69.4)
SI on T2WI			< 0.001*
hyperintensity	2(3.6)	15(17.7)
isointensity	6(10.9)	11(12.9)
slight hypointensity	29(52.7)	31(36.5)
mixed intensity	18(32.8)	28(32.9)
ADC value on DWI^#	0.892 ± 0.345	0.854 ± 0.212	0.985
Kinetic features
first-minute enhancement rate^#	154.807 ± 44.242	165.041 ± 49.677	0.389
maximum enhancement rate^#	1570.458 ± 494.581	1641.433 ± 609.262	0.650
Tmax^#	282.307 ± 131.737	265.302 ± 142.244	0.150
TIC			< 0.001*
persistent type	3(5.5)	5(5.9)
plateau type	28(50.9)	37(43.5)
washout type	24(43.6)	43(50.6)
Note: Independent t-test for continuous variables^# and Fisher’s-test or chi-square-test for classified variables
SI Signal intensity, T2WI T2-weighted image, ADC apparent diffusion coefficient, DWI diffusion-weighted imaging, TIC time-signal intensity curves,
Numbers in parentheses are percentages
^#Numbers are means ± standard deviations

Routine clinical MRI signature training and validation

We chose the features with significance at p < 0.05 during the first step. Then, using the six selected features with LR, MLP, RF, and SVM classifiers, the results showed that the training group AUCs were 0.65(CI:0.64–0.66), 0.99(0.99–0.99), 0.95(0.94–0.96), and 0.53(0.37–0.69), and those for the validation group were 0.59(0.46–0.72), 0.52(0.39–0.65), 0.53(0.46–0.60), and 0.49(0.38–0.60) (Table 2).

Table 2

Predictive Performance of the Routine Clinical MRI Signature for Ki-67 Expression in Luminal Breast Cancer
	LR	MLP	RF	SVM
AUC of training set	0.65(0.64–0.66)	0.99(0.99–0.99)	0.95(0.94–0.96)	0.53(0.37–0.69)
AUC of validation set	0.59(0.46–0.72)	0.52(0.39–0.65)	0.53(0.46–0.60)	0.49(0.38–0.60)
sensitivity	0.85	0.67	0.74	0.89
specificity	0.22	0.36	0.27	0.11
accuracy	0.60	0.55	0.56	0.59
Note: LR Logistic Regression, MLP Multilayer Perceptron, RF Random Forest, SVM Support Vector Machine, AUC area under the curve
Numbers in parentheses are 95% confidence intervals

Radiomic feature reliability and selection

Before radiomic analysis, a random cohort of 30 patients was chosen to assess intra- and interobserver reproducibility. The reproducibility was good, with an intraclass correlation coefficient (ICC) > 0.75. In the first step, significant radiomic features were selected from T2WI, DCE, DWI, and ADC sequences among patients in the training set. Specifically, 15, 3, 5, and 4 features were selected from each sequence, which showed apparent differences between the two groups. In the next step, 11 mpMRI radiomic features, incorporating the radiomic features from all four sequences, were finally selected. The selected mpMRI radiomic features are listed in Appendix Table 1.

Evaluation of machine learning classifiers

To build a prediction model for Ki-67 expression, we used algorithms from four machine learning models to establish, train, and validate single and multivariate models. Fivefold cross-validation was applied in the inner loop to ensure that these algorithms were adequately trained. The T2- and mpMRI radiomics-based models performed better than the DCE-, DWI-, and ADC-based models (all p < 0.05). Specifically, in the training sets, the four classifiers (LR, RF, MLP, SVM) achieved AUCs of 0.94 (0.92–0.96), 0.92 (0.90–0.94), 0.94 (0.92–0.96), and 0.91 (0.88–0.94), respectively, in the T2-based model and AUCs of 0.92 (0.90–0.94 (0.92–0.96), 0.90 (0.86–0.94), and 0.92 (0.89–0.95), respectively, in the mpMRI radiomic-based model. In the validation sets, the four classifiers yielded AUCs of 0.87 (0.85–0.89), 0.92 (0.88–0.96), 0.92 (0.84-1.00), and 0.89 (0.87–0.91) in the T2-based model and AUCs of 0.92 (0.88–0.96), 0.92 (0.90–0.94), 0.89 (0.85–0.93), and 0.92 (0.90–0.94) in the mpMRI radiomic-based model. Details of the prediction models' performance, including sensitivity, specificity, and accuracy, are depicted in Table 3 and can be observed in Fig. 3 using the graphs of ROC.

Table 3

Prediction Performance of the Radiomics Model for Ki-67 Expression in Luminal Breast Cancer
		T2WI	DCE	DWI	ADC	mpMRI
LR	AUC of training set	0.94(0.92–0.96)	0.63(0.60–0.66)	0.64(0.62–0.66)	0.68(0.65–0.71)	0.92(0.90–0.94)
	AUC of validation set	0.87(0.85–0.89)	0.67(0.65–0.69)	0.64(0.58–0.70)	0.64(0.56–0.70)	0.92(0.88–0.96)
	sensitivity	0.78	0.89	0.88	0.78	0.78
	specificity	0.90	0.23	0.31	0.26	0.95
	accuracy	0.83	0.64	0.67	0.58	0.84
MLP	AUC of training set	0.94(0.92–0.96)	0.62(0.57–0.67)	0.65(0.57–0.73)	0.67(0.63–0.71)	0.90(0.86–0.94)
	AUC of validation set	0.92(0.84-1.00)	0.62(0.54–0.70)	0.61(0.47–0.75)	0.64(0.50–0.78)	0.89(0.85–0.93)
	sensitivity	0.80	0.92	0.83	0.82	0.82
	specificity	0.91	0.14	0.29	0.24	0.89
	accuracy	0.84	0.62	0.62	0.60	0.85
RF	AUC of training set	0.92(0.90–0.94)	0.72(0.70–0.74)	0.73(0.71–0.75)	0.73(0.70–0.76)	0.94(0.92–0.96)
	AUC of validation set	0.92(0.88–0.96)	0.63(0.59–0.67)	0.63(0.61–0.65)	0.62(0.60–0.64)	0.92(0.90–0.94)
	sensitivity	0.73	0.76	0.92	0.82	0.86
	specificity	0.96	0.36	0.21	0.31	0.85
	accuracy	0.82	0.61	0.65	0.62	0.86
SVM	AUC of training set	0.91(0.88–0.94)	0.68(0.55–0.81)	0.58(0.45–0.71)	0.65(0.51–0.79)	0.92(0.89–0.95)
	AUC of validation set	0.89(0.87–0.91)	0.57(0.47–0.67)	0.64(0.36–0.92)	0.67(0.57–0.77)	0.92(0.90–0.94)
	sensitivity	0.70	0.94	1.00	0.91	0.82
	specificity	1.00	0.08	0.00	0.16	0.95
	accuracy	0.81	0.61	0.62	0.62	0.87
Note: LR Logistic Regression, MLP Multilayer Perceptron, RF Random Forest, SVM Support Vector Machine, T2WI T2-weighted image, DCE dynamic contrast enhancement, ADC apparent diffusion coefficient, DWI diffusion-weighted imaging, AUC area under the curve
Numbers in parentheses are 95% confidence intervals

The DeLong test results revealed the following: when comparing the pairwise AUC values of each model, the T2-based model and the mpMRI radiomic-based model exhibited significantly higher AUC values than the DWI-based, ADC-based, and DCE-based models (p < 0.05). However, when comparing the AUC of the T2- and mpMRI radiomics-based models, the results indicated no significant difference in their performance (p > 0.05) (Table 4).

Table 4

Delong Test Results for the AUC of Each Model
	LR		MLP		RF		SVM
	Z	P	Z	P	Z	P	Z	P
ADC-DCE	-0.351	0.726	0.188	0.851	-0.078	0.938	0.768	0.443
ADC-DWI	0.131	0.895	0.279	0.780	-0.109	0.912	1.111	0.266
ADC-T2WI	-2.583	0.009	-3.015	0.002	-3.393	0.001	-2.203	0.028
ADC-mpMRI	-3.280	0.001	-2.657	0.008	-3.465	0.001	-2.589	0.009
DCE-DWI	0.619	0.536	0.089	0.928	-0.028	0.977	0.039	0.969
DCE-T2WI	-2.316	0.021	-2.753	0.006	-3.194	0.001	-2.969	0.003
DCE-mpMRI	-2.932	0.003	-2.488	0.013	-3.239	0.001	-3.389	0.000
DWI-T2WI	-2.477	0.013	-2.787	0.005	-2.896	0.004	-3.262	0.001
DWI-mpMRI	-3.047	0.002	-2.372	0.018	-2.866	0.004	-3.675	0.000
T2WI-mpMRI	-1.262	0.207	0.613	0.540	0.000	1.000	-1.119	0.263
Note: LR Logistic Regression, MLP Multilayer Perceptron, RF Random Forest, SVM Support Vector Machine, T2WI T2-weighted image, DCE dynamic contrast enhancement, ADC apparent diffusion coefficient, DWI diffusion-weighted imaging

Constructing the Clinical-Radiomic nomogram.

The predictive performance of the nomogram for Ki-67 expression in luminal breast cancer was as follows: the AUC, sensitivity, specificity, and accuracy were 0.93 (0.90–0.96), 0.85, 0.82, and 0.84, respectively, in the training group. Moreover, the validation group had an AUC of 0.92 (0.89–0.95), sensitivity of 0.88, specificity of 1, and accuracy of 0.93 (Fig. 4).

Model calibration and analysis of the decision curve

We observed that the calibration curve had a high level of correlation between the predicted and actual probabilities of the nomogram, indicating excellent calibration of this predictive model (Fig. 5). We assessed the nomogram, mpMRI radiomics model, and routine clinical imaging model in predicting Ki-67 levels in luminal breast carcinoma through decision curve analysis (DCA). The nomogram outperformed the other two models, yielding superior net benefits (Fig. 6).

Our study revealed that machine learning models based on T2 radiomic and mpMRI radiomic signatures demonstrated significant potential to evaluate Ki-67 levels in luminal breast carcinoma under preoperative conditions. A clinical-radiomic nomogram model provided a visualized and individual method to preoperatively predict the status of Ki-67 and showed good predictive performance, with training and validation set AUCs of 0.93 and 0.92, respectively.

Ki-67 status could help to distinguish luminal breast cancer subtypes A and B (HER2 negative) and determine patients’ treatment selection and decision-making in clinical practice. Previous studies[19, 26, 27] have predicted Ki-67 expression by analyzing morphological features, ADC values, and hemodynamic parameters in conventional MR images, and the results demonstrated that some traditional MRI features are correlated with Ki-67 expression, which is consistent with the findings of this study. The present study showed that various parameters, including tumor maximum diameter, irregular margins, lobulated shape, invasion of adjacent structures, T2 signal, and differences in TICs between different categories, were statistically substantial (p < 0.05) and could be used to establish a prediction model. The MLP and RF models had high AUCs in the training sets, reaching 0.99 and 0.95, respectively. However, the predictive performance was poor in the validation sets (AUC < 0.6 for both models). Therefore, the predictive evaluation of clinical data and traditional MRI observations in determining the condition of Ki-67 in luminal breast cancer could be limited. Hence, a precise and reliable methodology that can clinically determine the Ki-67 propagation level before surgery is needed.

Radiomics is a promising approach that can noninvasively reflect the overall heterogeneity of a tumor by high-throughput extraction of potential data information within images. It can extract hidden structural information from conventional images and reduce the influence of personal bias. Previously, radiomics has helped predict the Ki-67 status in breast cancer, with inconsistent results. Liang et al. [20] showed that the radiomics model of T2WI could predict Ki-67 expression, possessing an AUC of 0.762, while DCE-MRI radiomic features were not significantly correlated with Ki-67 expression. Lafci et al[28] predicted Ki-67 expression in 73 cases of luminal breast cancer with an AUC of 0.746 in the validation set DCE model, with a sensitivity of 0.744 and a specificity of 0.643. Other studies [21, 22] revealed a significant association between radiomic characteristics depending on DCE-MR images and Ki-67 expression. However, the predictive performances of these studies were found to be suboptimal, with little clinical support for decision-making. We observed that the T2-based radiomics and mpMRI radiomics models demonstrated significant potential to forecast the Ki-67 levels in luminal breast cancer with AUCs > 0.87 for all four machine-learning models in the validation sets. The importance of T2WI helps determine diseased tissue by correlating diseased tissue with elevated free water, manifesting as regions of hyperintensity[29]. Our study is consistent with previous studies indicating a significant association between high SI on T2WI and mean Ki-67[24]. Moreover, the T2WI image-based radiomic classifier could substantially predict the Ki-67 status in breast carcinoma patients[20]. Breast cancers with significant propagation (i.e., high Ki-67) may exceed vascular oxygen allocation. This could extend hypoxia and cause necrosis, reducing cellular density[30]. Moreover, mpMRI can provide additional knowledge concerning tumor heterogeneity [31] and improve the predictive function of the model[23].

This paper combined radiomic features and four classifiers of advanced machine learning (LR, MLP, RF, SVM). Few previous studies have performed such a detailed predictive analysis to determine the Ki-67 levels in luminal breast cancer, with AUCs between 0.91 and 0.94 for the training group and between 0.87 and 0.92 for the validation group. The LR and RF classifiers exhibited excellent predictive performance, but the RF classifier showed overfitting to the training data, while the LR classifier had a stable performance in mpMRI radiomic training and validation sets, with AUC values of 0.92 for both groups. In our study, the other three types of classifiers yielded relatively satisfactory performance. The results indicated that machine learning with MRI-based radiomics models might noninvasively predict Ki-67 status and reduce unnecessary biopsy for breast carcinoma patients.

The clinical-radiomic nomogram model provided a slightly better predictive evaluation than the mpMRI radiomics model (AUC, 0.920 vs. 0.931) in the training group and a comparable performance (AUC, 0.920 vs. 0.921) in the validation group. Our study demonstrates a comparable trend to the findings of prior research[32] while showcasing superior predictive performance. The AUC value of 0.92 from our analysis significantly outperformed previous reports (AUC 0.793). The nomogram incorporating the Rad-score and clinical MRI variables can provide easy-to-use and individual predictions of the preoperative status of Ki-67 and shows significant efficacy in enhancing the predictive capabilities of the Ki-67 values in luminal breast cancer. Valuable calibration validated the nomogram model in the training and validation sets, providing evidence for better clinical management and useful clinical decision-making guidance.

Our study has certain limitations that should be noted. First, its retrospective design and the smaller sample size could have caused bias. Second, as a single-center study, with model performance evaluated using cross-validation, the lack of external validation from multiple centers is also a limitation. Therefore, the next step should include external model validation from centers with significantly more significant populations to validate the current findings.

Our study found that the T2-based radiomics model and mpMRI radiomics model combined with advanced machine learning classifiers could effectively predict the Ki-67 status in luminal breast cancer. The clinical-radiomic nomogram can provide an easy-to-use visual depiction and individual prediction of the preoperative status of Ki-67 and achieved a good performance. Our model could serve as an efficient tool to assist clinicians in personalized treatment for patients with luminal breast cancer.

Acknowledgments

Potential conflict of interest

The authors declare that they have no competing interests

Ethical approval

Institutional Review Board approval was obtained. (No.2018-L-23)

Informed consent

Informed consent was obtained from all individual participants included in the study.

Funding

This study has received funding by Priority Union Foundation of Yunnan Provincial Science and Technology Department and Kunming Medical University [No.202101AY070001-101]

Authors' contributions

Conceptualization: Qian Gao, Meixiu Lu, JunLi

Data curation: Yanhui Lu, Ruixue Zhao, Linlin Fang

Funding acquisition: JunLi

Investigation: Meixiu Lu, Xiaojie Xie

Methodology: Qian Gao, Meixiu Lu, JunLi

Project administration: Qian Gao, Meixiu Lu, JunLi

Resources: Xiaojie Xie, Chunyan Luo, Chao Gao, Zhiquan Han

Supervision: Dan Han, JunLi

Visualization: Qian Gao, Meixiu Lu

Writing-original draft: Qian Gao, Meixiu Lu

Writing-review & editing: Qian Gao, Meixiu Lu, JunLi

All authors read and approved the final manuscript

Harbeck N, Gnant M. Breast cancer. Lancet. 2017; 389: 1134-50.
Goldhirsch A, Wood WC, Coates AS, Gelber RD, Thürlimann B, Senn HJ. Strategies for subtypes--dealing with the diversity of breast cancer: highlights of the St. Gallen International Expert Consensus on the Primary Therapy of Early Breast Cancer 2011. Ann Oncol. 2011; 22: 1736-47.
Husni Cangara M, Miskad UA, Masadah R, Nelwan BJ, Wahid S. Gata-3 and KI-67 expression in correlation with molecular subtypes of breast cancer. Breast Dis. 2021; 40: S27-s31.
Chen Y, Liu X, Yu K, Sun X, Xu S, Qiu P, et al. Impact of hormone receptor, HER2, and Ki-67 status conversions on survival after neoadjuvant chemotherapy in breast cancer patients: a retrospective study. Ann Transl Med. 2022; 10: 93.
Kanyılmaz G, Yavuz BB, Aktan M, Karaağaç M, Uyar M, Fındık S. Prognostic Importance of Ki-67 in Breast Cancer and Its Relationship with Other Prognostic Factors. Eur J Breast Health. 2019; 15: 256-61.
Goldhirsch A, Winer EP, Coates AS, Gelber RD, Piccart-Gebhart M, Thürlimann B, et al. Personalizing the treatment of women with early breast cancer: highlights of the St Gallen International Expert Consensus on the Primary Therapy of Early Breast Cancer 2013. Ann Oncol. 2013; 24: 2206-23.
Inwald EC, Koller M, Klinkhammer-Schalke M, Zeman F, Hofstädter F, Gerstenhauer M, et al. 4-IHC classification of breast cancer subtypes in a large cohort of a clinical cancer registry: use in clinical routine for therapeutic decisions and its effect on survival. Breast Cancer Res Treat. 2015; 153: 647-58.
Harbeck N, Thomssen C, Gnant M. St. Gallen 2013: brief preliminary summary of the consensus discussion. Breast Care (Basel). 2013; 8: 102-9.
Burstein HJ, Curigliano G, Loibl S, Dubsky P, Gnant M, Poortmans P, et al. Estimating the benefits of therapy for early-stage breast cancer: the St. Gallen International Consensus Guidelines for the primary therapy of early breast cancer 2019. Ann Oncol. 2019; 30: 1541-57.
Brown J, Scardo S, Method M, Schlauch D, Misch A, Picard S, et al. A real-world retrospective study of the use of Ki-67 testing and treatment patterns in patients with HR+, HER2- early breast cancer in the United States. BMC Cancer. 2022; 22: 502.
Davey MG, Hynes SO, Kerin MJ, Miller N, Lowery AJ. Ki-67 as a Prognostic Biomarker in Invasive Breast Cancer. Cancers (Basel). 2021; 13.
Li L, Han D, Wang X, Wang Q, Tian J, Yao J, et al. Prognostic values of Ki-67 in neoadjuvant setting for breast cancer: a systematic review and meta-analysis. Future Oncol. 2017; 13: 1021-34.
Kalvala J, Parks RM, Green AR, Cheung KL. Concordance between core needle biopsy and surgical excision specimens for Ki-67 in breast cancer - a systematic review of the literature. Histopathology. 2022; 80: 468-84.
Ahn S, Lee J, Cho MS, Park S, Sung SH. Evaluation of Ki-67 Index in Core Needle Biopsies and Matched Breast Cancer Surgical Specimens. Arch Pathol Lab Med. 2018; 142: 364-8.
Fan M, Yuan W, Zhao W, Xu M, Wang S, Gao X, et al. Joint Prediction of Breast Cancer Histological Grade and Ki-67 Expression Level Based on DCE-MRI and DWI Radiomics. IEEE J Biomed Health Inform. 2020; 24: 1632-42.
Li C, Song L, Yin J. Intratumoral and Peritumoral Radiomics Based on Functional Parametric Maps from Breast DCE-MRI for Prediction of HER-2 and Ki-67 Status. J Magn Reson Imaging. 2021; 54: 703-14.
Balachandran VP, Gonen M, Smith JJ, DeMatteo RP. Nomograms in oncology: more than meets the eye. Lancet Oncol. 2015; 16: e173-80.
Vickers AJ, Elkin EB. Decision curve analysis: a novel method for evaluating prediction models. Med Decis Making. 2006; 26: 565-74.
Mori N, Ota H, Mugikura S, Takasawa C, Ishida T, Watanabe G, et al. Luminal-type breast cancer: correlation of apparent diffusion coefficients with the Ki-67 labeling index. Radiology. 2015; 274: 66-73.
Liang C, Cheng Z, Huang Y, He L, Chen X, Ma Z, et al. An MRI-based Radiomics Classifier for Preoperative Prediction of Ki-67 Status in Breast Cancer. Acad Radiol. 2018; 25: 1111-7.
Juan MW, Yu J, Peng GX, Jun LJ, Feng SP, Fang LP. Correlation between DCE-MRI radiomics features and Ki-67 expression in invasive breast cancer. Oncol Lett. 2018; 16: 5084-90.
Qiu X, Wang H, Wang Z, Fu Y, Yin J. Expression of PCNA, Ki-67 and COX-2 in breast cancer based on DCE-MRI image information. J Infect Public Health. 2020; 13: 2032-7.
Liu W, Cheng Y, Liu Z, Liu C, Cattell R, Xie X, et al. Preoperative Prediction of Ki-67 Status in Breast Cancer with Multiparametric MRI Using Transfer Learning. Acad Radiol. 2021; 28: e44-e53.
Song SE, Cho KR, Cho Y, Kim K, Jung SP, Seo BK, et al. Machine learning with multiparametric breast MRI for prediction of Ki-67 and histologic grade in early-stage luminal breast cancer. Eur Radiol. 2022; 32: 853-63.
Fedorov A, Beichel R, Kalpathy-Cramer J, Finet J, Fillion-Robin JC, Pujol S, et al. 3D Slicer as an image computing platform for the Quantitative Imaging Network. Magn Reson Imaging. 2012; 30: 1323-41.
Yoon GY, Choi WJ, Cha JH, Shin HJ, Chae EY, Kim HH. The role of MRI and clinicopathologic features in predicting the invasive component of biopsy-confirmed ductal carcinoma in situ. BMC Med Imaging. 2020; 20: 95.
Caiazzo C, Di Micco R, Esposito E, Sollazzo V, Cervotti M, Varelli C, et al. The role of MRI in predicting Ki-67 in breast cancer: preliminary results from a prospective study. Tumori. 2018; 104: 438-43.
Lafcı O, Celepli P, Seher Öztekin P, Koşar PN. DCE-MRI Radiomics Analysis in Differentiating Luminal A and Luminal B Breast Cancer Molecular Subtypes. Acad Radiol. 2023; 30: 22-9.
Santamaría G, Velasco M, Bargalló X, Caparrós X, Farrús B, Luis Fernández P. Radiologic and pathologic findings in breast tumors with high signal intensity on T2-weighted MR images. Radiographics. 2010; 30: 533-48.
Leek RD, Landers RJ, Harris AL, Lewis CE. Necrosis correlates with high vascular density and focal macrophage infiltration in invasive carcinoma of the breast. Br J Cancer. 1999; 79: 991-5.
Bi S, Li J, Wang T, Man F, Zhang P, Hou F, et al. Multi-parametric MRI-based radiomics signature for preoperative prediction of Ki-67 proliferation status in sinonasal malignancies: a two-centre study. Eur Radiol. 2022; 32: 6933-42.
Wu J, Fang Q, Yao J, Ge L, Hu L, Wang Z, et al. Integration of ultrasound radiomics features and clinical factors: A nomogram model for identifying the Ki-67 status in patients with breast carcinoma. Front Oncol. 2022; 12: 979358.

appedendix.docx

Download PDF

Version 1

posted

You are reading this latest preprint version

Machine Learning with Multiparametric MRI-based Radiomics Models for Preoperative Prediction of Ki-67 Status in Luminal Breast Cancer

Status:

Version 1

Abstract

Figures

Introduction

Materials and Methods

Patient enrollment

MRI examination

Routine clinical MRI evaluation

Radiomic feature extraction

Selection of characteristic features and development of radiomic signatures

Establishment and evaluation of machine learning classifiers

Constructing the clinical-radiomic nomogram and decision curve analysis

Statistical analysis

Results

Routine clinical MRI signature

Routine clinical MRI signature training and validation

Radiomic feature reliability and selection

Evaluation of machine learning classifiers

Model calibration and analysis of the decision curve

Discussion

Declarations

References

Supplementary Files

Status:

Version 1