Robustness of radiomic features of benign breast lesions and hormone receptor positive/HER2-negative cancers across DCE-MR magnet strengths

Abstract

Radiomic features extracted from breast lesion images have shown potential in diagnosis and prognosis of breast cancer. As medical centers transition from 1.5 T to 3.0 T magnetic resonance (MR) imaging, it is beneficial to identify potentially robust radiomic features across field strengths because images acquired at different field strengths could be used in machine learning models. Dynamic contrast-enhanced MR images of benign breast lesions and hormone receptor positive/HER2-negative (HR+/HER2−)breast cancers were acquired retrospectively, yielding 612 unique cases: 150 and 99 benign lesions imaged at 1.5 T and 3.0 T, and 223 and 140 HR+/HER2− cancerous lesions imaged at 1.5 T and 3.0 T, respectively. In addition, an independent set of seven lesions imaged at both field strengths, three benign lesions and four HR+/HER2− cancers, was analyzed separately. Lesions were automatically segmented using a 4D fuzzy c-means method; thirty-eight radiomic features were extracted. Feature value distributions were compared by cancer status and imaging field strength using the Kolmogorov-Smirnov test. Features that did not demonstrate a statistically significant difference were considered to be potentially robust. The area under the receiver operating characteristic curve (AUC), for the task of classifying lesions as benign or HR+/HER2− cancer, was determined for each feature at each field strength. Three features were found to be both potentially robust across field strength and of high classification performance, i.e., AUCs statistically greater than 0.5 in the classification task: one shape feature (irregularity), one texture feature (sum average) and one enhancement variance kinetics features (enhancement variance increasing rate). In the demonstration set of lesions imaged at both field strengths, two of the three potentially robust features showed qualitative agreement across field strength. These findings may contribute to the development of computer-aided diagnosis models that are robust across field strength for this classification task.

1. Introduction

Radiomic features extracted from magnetic resonance (MR) images are being investigated for use in diagnosis and prognosis of breast cancer [1-7]. Clinical breast dynamic contrast-enhanced (DCE) MR imaging is typically performed at 1.5 T or 3.0 T, depending upon institutional resources and scheduling. While a recent study failed to demonstrate significant difference in diagnostic accuracy when physicians utilized images acquired at 1.5 T or 3.0 T for breast cancer diagnosis [8], there are other benefits for choosing one field strength for imaging over the other, such as lower cost (in the case of lower field strength) or the potential for reduction of contrast agent used (in the case of higher field strength).

Radiomic features extracted from DCE-MRI can characterize several different aspects of lesions, including their size, shape, morphology, texture, the kinetics of contrast enhancement, and enhancement variance [9-13]. All other factors being equal, features related to size, shape, and morphology should not depend upon image acquisition field strength. Some features related to kinetic curve assessment (such as the maximum enhancement) and enhancement variance kinetics may depend upon field strength, as relative signal enhancement is greater for higher field strength because of increases in T1, the longitudinal relaxation time [14-16]. Water exchange effects also vary with field strength, which can affect T1 [17]. Cancers typically contain more iron, including iron bound in ferritin [18], than do benign tissues. This may have a differential effect on extracted features at different field strengths because of the difference in relaxivity and magnetic susceptibility gradients between tissues and different field strengths [19,20]. Additionally, if not explicitly controlled for, MR images acquired at different field strengths may have different image acquisition parameters not inherently tied to field strength, such as spatial resolution, and this could affect texture features. Image acquisition factors such as the number of elements in a coil may differ between scanners at different field strengths and affect the signal to noise ratio. These variations together, some inherent to field strength alone and some not, may influence the calculation of some radiomic features. Therefore, there is a need to assess how radiomic features perform across imaging protocols [21,22].

This study aims to identify features extracted from DCE-MR images of beast lesions that may be both robust across field strength and useful for diagnostic classification, and to assess the impact of field strength on the classification performance of radiomic features. The study is focused on cancers of a specific molecular subtype (hormone receptor (HR) positive/human epidermal growth factor receptor 2 (HER2) negative, i.e., HR+/HER2−) because a previous study has indicated that radiomic feature distributions differ by molecular subtype [23]. We investigated feature robustness and classification performance across field strength considering several confounding variables, such as spatial resolution, using two datasets: (1) a retrospectively-collected large clinical dataset of subjects who underwent DCE-MR imaging at either 1.5 T or 3.0 T and (2) an additional retrospective small set of cases from patients who had undergone MR imaging of the breast at both field strengths.

2. Materials and Methods

2.1. Investigation of field strength dependence of radiomic features extracted from 612 lesions imaged at either 1.5 T or 3.0 T field strength.

2.1.1. Description of the 612-lesion database

Our study included DCE-MR images of 612 unique breast lesions acquired at either 1.5 T or 3.0 T during the time period of 2005-2016 (Table 1) and collected retrospectively under HIPAA/IRB compliance. There was one lesion per case. Criteria for inclusion in the study were that the field strength of imaging was known and, for cancer cases, that immunohistochemical staining indicated that they were positive for estrogen receptor, negative or positive for progesterone receptor, and negative for human epidermal growth factor receptor 2 (HR+/HER2−) [24]. All lesions were imaged in the axial plane using Philips Achieva or Philips Intera Achieva scanners with a T1-weighted spoiled gradient sequence. The repetition time across all lesions was (median, [95 % confidence interval (CI)]) 5.48 [4.70, 7.22] ms, while the echo time was 2.72 [2.33, 3.60] ms. The flip angle for most lesions imaged at 1.5 T was 12 degrees, while the flip angle for most lesions imaged at 3.0 T was 10 degrees. The difference in flip angles was compensated for the differences in average T1 at the two field strengths The field of view across all lesions was 360 [300, 400] mm. The temporal resolution of the DCE-MR imaging series across all lesions was 65.2 [54.4, 84.6] s. Images were not acquired using a single scanner or with a single prospective protocol, a reflection of the clinical basis of the image acquisitions and the retrospective nature of the study. Because the study was designed to investigate robustness of features in the context of image acquisition, no resampling of the images, correction for magnetic field inhomogeneities, or data harmonization was conducted. In this 612-lesion database, lesions were imaged at one or the other field strength.

Table 1:

Description of the main analysis database of 612 unique lesions (benign lesions and hormone receptor positive/HER2 negative (HR+/HER2−) cancers imaged at either 1.5 T or 3.0 T): number of cases, patient age, TNM staging size, and maximum lesion size by lesion type and field strength of the image acquisition (CI = confidence interval). TNM staging size is from the primary tumor, lymph node, and metastasis (TNM) classification of the American Joint Commission of Cancer (AJCC) [25].

	Number of lesions at field strength of image acquistion			Patient age* (Median [95% CI], years)		maximum lesion size (radiomic feature S4) (Median, [95% CI], mm)		TNM staging size Number of lesions (% of lesion type by field strength)
Lesion type	1.5 T	3.0 T	Total	1.5 T	3.0 T	1.5 T	3.0 T		1.5 T	3.0 T
Benign	150	99	249	49 [25, 73]	47 [32, 74]	12.7 [5.3, 56.3]	13.0 [5.5, 54.1]	T1 (≤ 20 mm)	112 (75%)	72 (73%)
								T2 (20 mm ≥ 50 mm	32 (21%)	24 (24%)
								T3 (> 50 mm)	6 (4%)	3 (3%)
HR+ / HER2−	223	140	363	58 [35, 83]	55 [38, 80]	24.9 [8.3, 96.4]	22.6 [5.2, 87.4]	T1 (≤ 20 mm)	83 (37%)	63 (45%)
								T2 (20 mm ≥ 50 mm	105 (47%)	61 (44%)
								T3 (> 50 mm)	35 (16%)	16 (11%)
Total	373	239	612

^*Patient age was not available for 33 out of 150 benign lesions imaged at 1.5 T (22%), 13 out of 99 benign lesions imaged at 3.0 T (13%), 22 out of 223 HR+/HER2−cancers imaged at 1.5 T (10%), and 10 out of 140 HR+/HER2−cancers imaged at 3.0 T (7%).

The distribution of voxel sizes for each lesion type by field strength, for lesions imaged at either 1.5 T or 3.0 T, spanned 1-1.5 mm³ for most images acquired at 1.5 T, and 0.5-1 mm³ for most images acquired at 3.0 T. Slice thickness varied from 1.6 to 2.5 mm.

2.1.2. Lesion segmentation

Lesions were segmented using a 4D fuzzy c-means (FCM) method requiring only the manual indication of a seed-point [26]. The FCM method is an unsupervised learning technique in which the algorithm uses the kinetic curves, i.e., the multiple voxel values over time (normalized to the pre-contrast image) to partition the voxels in the image space into clusters by using a fuzzy partition yielding the degree of membership that continuously ranges from 0 to 1. The voxels are then automatically separated into lesion and non-lesion classes. For each lesion, segmentation was additionally verified by an experienced breast imaging analyst after it was completed, via visual inspection. Lesion segmentation was conducted using Matlab R2018b (MathWorks; Natwick, MA) and software written in-house.

2.1.3. Radiomic feature extraction

After lesion segmentation was conducted, feature extraction process was entirely automatic. Thirty-eight radiomic features were extracted from the 3D lesions [10-13]. These features are comprised of 4 size features, 3 shape features, 3 morphology features, 14 texture features, 10 kinetic curve assessment features, and 4 enhancement variance kinetics features (a full list of radiomic features extracted from the images, along with descriptions and their abbreviations used in this work, is given in Appendix A.) These 38 features were chosen due to previous studies indicating their usefulness for this classification task and their comprehensive coverage of the range of malignant and benign lesion characteristics [2,13], as well as to reduce the potential for overfitting that can exist when a high number of features are used (Hua et al., 2005). Since no lesions in this dataset were imaged at both field strengths, features related to size (4 features) were not used in the analysis; thirty-four features remained for the analysis. This is because this population-based study was limited to considering the characteristics of a lesion within a population, as opposed to the dimensions of individual lesions, which could be a function of when the lesion was imaged during its growth [27]. To calculate the kinetic-related radiomic features, all time point images were used. A 32 binned co-occurrence matrix was used on the first post-contrast image to calculate the 3D texture features. No variation in bin size was investigated in this study. Feature extraction was conducted using Matlab R2018b (MathWorks; Natwick, MA) and software written in-house.

2.1.4. Statistical analysis

The age of subjects for each lesion type (Table 1) across field strength was compared using the Kolmogorov-Smirnov test [28,29]. To do so, data imputation to address missing cases was conducted by using a moving median of 15 neighboring cases for each missing case. The distribution of lesion sizes for each lesion type was compared between field strength of imaging acquisition using the Kolmogorov-Smirnov test, with the null hypothesis that the distributions were drawn from the same population. The distribution of lesions by primary tumor, lymph node, and metastasis (TNM) classification of the American Joint Commission of Cancer [25] was compared for each lesion type across field strength of imaging using Pearson’s chi-squared test. Each comparison was considered statistically significant if P < 0.05.

Because field strength of imaging is associated with difference in spatial resolution due to differences in scanning protocols, the voxel size distributions (i.e., the distributions of voxel sizes from the cases within each field strength) were compared across field strength using the Kolmogorov-Smirnov test, with the null hypothesis that the distributions were drawn from the same population. The comparison was deemed statistically significant if P < 0.05.

For each of the thirty-four radiomic features evaluated in the study, the distributions of feature values by lesion type and by field strength were compared using the Kolmogorov-Smirnov test. Features were considered potentially robust when they failed to reject the null-hypothesis that their distributions were the same across field strengths, i.e., when their P-value was greater than the Bonferroni-Holm adjusted significance level [30]. The significance level was adjusted due to multiple comparisons (thirty-four), for both lesion types. Multiple comparisons correction was necessary to address potential issues with the false discovery rate, since thirty-four features were investigated.

The performance of each feature in the task of classifying lesions as benign or HR+/HER2− was assessed at each field strength by the area under the receiver operating characteristic (ROC) curve (AUC) [31] as performance metric using the conventional binormal ROC model [32]. The median AUC for each feature at each field strength and its 95% CI were determined with 2000 bootstrap iterations.

For features that were potentially robust with respect to field strength and performed greater than chance in the task of classifying lesions as benign or HR+/HER2−, correlation of features with voxel size was assessed for lesions by field strength of acquisition and by lesion status using the Pearson correlation coefficient [33]. Features were considered to be correlated with voxel size when the P-value was less than the Bonferroni-Holm-adjusted significance level, which was adjusted from P = 0.05 due to multiple comparisons (thirty-four).

Statistical analysis was conducted using Matlab R2019b (MathWorks; Natwick, MA).

2.2. Comparison of radiomic features extracted from a demonstration 7-lesion dataset of subjects imaged at both 1.5 T and 3.0 T.

A smaller demonstration study was conducted using features extracted from DCE-MR images from seven cases obtained at both field strengths between August 2011 and May 2012 (Table 2). There was one lesion per case. These subjects had undergone DCE-MR breast imaging at field strength of 1.5 T and 3.0 T using Philips Achieva scanners with consistent coil geometry, as part of another study [34]. The criterion for inclusion in our current study was that the lesions were benign or cancers of molecular subtype HR+/HER2−, as confirmed through pathology using the same criteria used for the larger dataset. The temporal resolution of the DCE-MR imaging series across all lesions (mean, [range]) was 76.3 [74.9, 79.0] s. All lesions were imaged post-biopsy and, for the cancerous lesions, before any treatment. The time elapsed between imaging at both field strengths ranged from 1 to 90 days. Similar to the first study, no image resampling or magnetic field inhomogeneity correction was conducted for these images.

Table 2:

Description of the demonstration database (7 cases imaged at both field strengths): lesion type, BI-RADS designation, age at imaging, time between imaging at the two field strengths, maximum lesion diameter (radiomic feature S4) as extracted from imaging at each field strength, and voxel resolution at each field strength. As available from radiological reports, additional information is given in the lesion type field. ME represents mass-enhancement, while NMLE stands for non-mass-like enhancement.

Subject number	Lesion Type	BI- RADS	Age at imaging (years)	Time between imaging at the two field strengths (days)	Maximum lesion diameter (radio mic feature S4) (mm)		Voxel resolution (given as x × y × z, mm)
					1.5 T	3.0 T	1.5 T	3.0 T
1	Benign	2	48	90	60.0	74.3	0.71 × 0.71 × 2	0.75 × 0.75 × 1.6
2	Benign (amorphous calcifications, NMLE)	4	61	20	76.7	75.0	0.75 × 0.75 × 2	0.75 × 0.75 × 1.6
3	Benign (fibroepithelial lesion)	4	40	1	15.0	13.2	0.76 × 0.76 × 1.6	0.75 × 0.75 × 1.6
4	HR+/HER2− (pleomorphic calcifications, ME within NMLE)	6	42	11	87.6	91.9	0.75 × 0.75 × 2	0.75 × 0.75 × 1.6
5	HR+/HER2− (ME)	6	51	16	20.5	18.6	0.74 × 0.74 × 2	0.74 × 0.74 × 1.6
6	HR+/HER2− (ME)	6	35	2	24.3	21.2	0.76 × 0.76 × 2	0.60 × 0.60 × 1.6
7	HR+/HER2− (NMLE)	6	60	7	108.0	110.8	0.76 × 0.76 × 2	0.74 × 0.74 × 1.6

Lesions were segmented and radiomic features were extracted using the same methods described above for the larger dataset, for the specific purpose of using the paired cases to (1) verify that features expected to be independent of field strength alone were indeed so, and (2) observe the feature values of features potentially dependent on field strength but found to be robust in the first study. No image registration was conducted in the comparison of radiomic features of lesions imaged at the two magnet strengths. All but one of these cases showed greater than 10% difference in voxel size, with subject number 3 being the exception. Kinetic curves representing the signal from the mean of all voxels within each lesion and the signal from the most enhancing voxels were collected for qualitative comparison between field strengths [12].

Bland-Altman figures [35] were used to assess the difference in feature value (as ratio of difference to the average of feature value, due to the range of orders of magnitude) as a function of the average of feature value for the features identified as both potentially robust and useful for classification from the first study. Additionally, size, shape, and morphology features, which should not vary by field strength of acquisition, were also evaluated using Bland-Altman figures for comparison.

3. Results

3.1. Investigation of field strength dependence of radiomic features extracted from 612 lesions imaged at either 1.5 T or 3.0 T field strength.

Distribution of patient age across field strength demonstrated significant difference (P =0.04) for benign lesions but not for luminal A cancers (P = 0.50). The difference in patient age for benign lesions across field strength could have been a function of the data imputation method, as the comparison did not show significant difference across field strength for the subjects for which age was available from the database (P = 0.34 for benign lesions, P = 0.44 for HR+/HER2− cancers). This effect may have been more prominent for the benign lesions due to the larger proportion of benign lesions for which age was not available, compared to the HR+/HER2− cancers. The distributions of lesion sizes for benign lesions and HR+/HER2− cancers failed to demonstrate significant difference across field strength (P = 0.98 and P = 0.59, respectively). The distribution of lesion sizes for benign lesions and HR+/HER2− cancers, determined by TNM staging, failed to demonstrate significant difference across field strength (P = 0.82 and P = 0.27, respectively).

The distributions of image voxel sizes for benign lesions and HR+/HER2− cancers were significantly different across field strength based on the Kolmogorov-Smirnov test (P < 0.0001 for each), indicating that voxel size difference could be a factor that influences potential robustness of the features with respect to field strength.

In the analysis of the feature distributions, P-values from the Kolmogorov-Smirnov test were compared to the significance level adjusted for multiple comparisons using the Bonferroni-Holm correction (thirty-four for each lesion type). This multiple comparison method identifies three categories of comparisons: (a) feature value distributions that failed to demonstrate significant difference across field strength (unadjusted P > 0.05), (b) feature value distributions for which the unadjusted P was less than the significance level (P < 0.05) but the comparison failed to demonstrate significant difference according to the Bonferroni-Holm-adjusted significance level, and (c) feature value distributions for which the difference remained significant after adjusting the significance level for multiple comparisons. From this analysis, five features were identified as potentially robust (i.e., in category (a) or (b) as noted above) across field strength: the shape feature of irregularity (G2), the texture feature of sum average (T11), and three out of four enhancement variance features (maximum enhancement variance (E1), enhancement variance increasing rate (E3), enhancement variance decreasing rate (E4) were potentially robust across field strength for both benign lesions and HR+/HER2− cancers (Figure 1).

Figure 1:

P-values from the Kolmogorov-Smirnov test for the null hypothesis that two groups being compared came from the same distribution, for extracted feature values for benign (top, green) and HR+/HER2− (bottom, red) lesions from DC-MR images acquired at 1.5 T and 3.0 T. Green symbols indicate values for benign lesions, and red symbols indicate values for HR+/HER2− lesions. A logarithmic scale is used to aid in visualization of the range of p-values. Circles indicate features for which feature value distributions failed to demonstrate significant difference across field strength (unadjusted p > 0.05) and stars indicate features for which the difference remained significant after adjusting the significance level for multiple comparisons. Another possible result could have been features for which the unadjusted p was less than 0.05 while the comparison failed to demonstrate significant difference according to the Bonferroni-Holm-adjusted significance level. However, this was not present for any feature. The gray shading (features S1-S4) indicates results for the size features, which were not included in the analysis (G: shape features, M: morphology features, T: enhancement texture features, K: kinetic curve assessment features, E: enhancement variance kinetics. The full set of abbreviations is available in Appendix A.) [total number of cases: 612 unique lesions]

The median AUC and 95% confidence interval of the AUC in the clinical task of distinguishing between benign lesions and HR+/HER2− breast cancers were calculated for each feature and, for potentially robust features, the values were compared for each field strength of image acquisition (Figure 2). Three of these potentially robust features, irregularity (G2), sum average (T11), and enhancement variance increasing rate (E3), were found to be useful for the task of classifying benign lesions and HR+/HER2− cancers, as the lower bound of the 95% confidence interval of the AUC for this task was greater than 0.5 for lesions at each field strength for these features.

Figure 2:

Median area under the receiver operating characteristic curve for features extracted from images at 3.0 T versus the AUC for features extracted from images at 1.5 T, in the task of distinguishing between benign lesions and HR+/HER2− breast cancers. Large circles (and error bars representing 95% confidence intervals) indicate potentially robust features (abbreviations as in Appendix A) and small circles (without error bars) indicate features which were not robust across field strength. [total number of cases: 612 unique lesions]

All three features that demonstrated both potential robustness with respect to field strength for both lesion types and performance better than chance in the task of classifying lesions as benign or HR+/HER2− breast cancers demonstrated potential robustness with respect to voxel size for all four combinations of lesion types and field strength of image acquisition, as determined by the Pearson correlation coefficient, since the P-value for each feature was greater than the Bonferroni-Holm-adjusted significance level.

Box plots show the median and range of feature values for the three features that demonstrated both potential robustness with respect to field strength for both lesion types and classification performance better than chance (Figure 3).

Figure 3:

Box plots for features potentially robust with respect to field strength with classification performance greater than chance. Because the feature values for enhancement-variance increasing rate (E3) included a wide range of values, the box plot for this feature is shown with a logarithmic scale. [total number of cases: 612 unique lesions]

3.2. Comparison of radiomic features extracted from a demonstration 7-lesion dataset of subjects imaged at both 1.5 T and 3.0 T.

For the small case study of subjects imaged at both field strengths, the kinetic curves for each lesion (Figure 4) show an overall consistency in their behavior across the two field strengths. The maximum enhancement was higher for the kinetic curves extracted from the most-enhancing voxels than averaged over the entire lesion as expected except for subject 6, a subject with a HR+/HER2− cancer. Images acquired at 1.5 T for this subject appear to have been affected by patient motion for the second and later post-contrast images, as indicated by the decrease in the normalized signal of the most enhancing voxels.

Figure 4:

Central slice subtraction (pre-contrast image subtracted from second post-contrast image) DCE-MR images and kinetic curves (normalized to the pre-contrast value) from the demonstration set of subjects imaged at both field strengths. Column 1: images acquired at 1.5 T; column 2: kinetic curves derived from all voxels within the lesion (open circles) and from the most enhancing voxels (closed circles) for lesions imaged at 1.5 T; column 3: images acquired at 3.0 T; column 4: kinetic curves for lesions imaged at 3.0 T. Red outlines indicate 3D segmentation as determined using fuzzy c-means.

Bland-Altman figures for features expected to be independent of field strength alone (categories of size, shape excluding irregularity, which is discussed below, and morphology) show that the most difference for feature values for the paired set was present in the morphology features.

When the three features identified from the first study as both potentially robust across field strength and useful in classifying lesions as benign or HR+/HER2− cancer (G2, T11, and E3; Figure 3) were compared across field strength using Bland-Altman figures for the small dataset of seven lesions, the differences in values between field strengths appeared substantially larger for the enhancement variance increasing rate feature (E3) than the variation in irregularity (G2) and sum average (T11) (Figure 6). These latter two features were comparable to variation in most of the size, shape, and morphology features.

Figure 6:

Figures demonstrating similarity in feature values across magnet strength. Bland-Altman figures are shown for features potentially robust with respect to field strength with classification performance greater than chance, as identified in the first study, for each lesion imaged at two separate field strengths in the second study. Green circles indicate benign lesions; red circles indicate HR+/HER2− lesions. The numbers in the circles correspond to the subject numbers in Table 2. Note that for one feature, enhancement-variance increasing rate (E3), the data are shown with a logarithmic scale on the x-axis.

4. Discussion

The question of variation of radiomic features with image acquisition has been the subject of study by others, as it is essential for ultimate clinical implementation of radiomics / computer-aided diagnosis into clinical workflow. There have been several studies across different imaging modalities that have investigated the association of texture features with imaging acquisition parameters such as slice thickness [36,37], spatial resolution and sampling bandwidth [38-40], and size of tissue architecture with respect to imaging resolution [37,39,40]. The effect of factors such as image resolution, slice thickness, number of detector elements, field strength, field inhomogeneity, and scanner manufacturer on kinetic curve features broadly has also been investigated [34,41-44]. These studies demonstrate the many factors that impact extracted radiomic features and the interest in characterizing radiomic features with respect to image acquisition.

Our study investigated the potential robustness of radiomic features in two collections of benign lesions and HR+/HER2− cancers of the breast imaged at two different field strengths, one in a large population of individual cases imaged at either field strength, and one in a small group of patients imaged at both field strengths. As noted by others, it may be useful to identify radiomic features that are robust across different imaging standards, of which field strength is one element, so that imaging protocols would not have to be tailored to feature extraction [45]. Collaborations such as the Image Biomarker Standardization Initiative [46] have also been interested in investigating robustness of radiomic features through efforts such as developing benchmark datasets and benchmark values. Our study been designed to contribute to understanding of features with respect to field strength of imaging and builds upon the efforts by others to understand the variation of radiomic features in the context of image acquisition parameters.

In our study of 612 unique lesions, the shape feature of irregularity (G2) was found to be potentially robust with respect to field strength and performed greater than chance in the classification task. While in theory there could be differences in the feature of irregularity as a function of field strength due to differences in variation in T1 or in contrast media relaxivity, these differences were not large enough to influence our analysis. In comparison, a published lesion signature for distinguishing between benign lesions and HR+/HER2− cancers identified irregularity as a dominant feature of that signature [4]. Together, these studies emphasize the utility of irregularity as a radiomic feature that is both robust and helpful in classification.

Previous work has also investigated the effect of biopsy on each radiomic feature using a similar analysis of feature value distribution and classification performance [47]. For the feature of irregularity (G2), benign lesions demonstrated significant difference in the feature value distribution between those imaged pre-biopsy and those imaged post-biopsy while HR+/HER2− lesions failed to do. Both lesion types failed to demonstrate significant difference in feature value for the texture feature sum average (T11) and enhancement variance increasing rate (E3) when compared across biopsy conditions, and these features performed better than chance in classification. In this study, we elected to not use only lesions known to be imaged pre-biopsy as this would have reduced statistical power, especially as only 43 cancers were known to be imaged pre-biopsy. Therefore, there is some possibility that biopsy condition acts as an additional confounding variable.

Together, these three studies highlight the usefulness and potential robustness of the quantitative feature of irregularity in the diagnostic classification of benign lesions and HR+/HER2− cancers relative to both field strength and biopsy condition.

Only one texture feature was found to be potentially robust with respect to field strength. This is not surprising given the significant difference in voxel size by field strength, with similar results having been found in computed tomography studies [45]. No kinetic curve features were shown to be potentially robust across field strength, which is not surprising because of the inherent expected dependence of these features upon field strength. It should additionally be considered that the difference in spatial resolution could have also affected their robustness, as previously seen [41].

In our demonstration assessment of the seven cases (three benign lesions and four HR+/HER2− cancers) imaged at both field strengths, of the potentially robust features, only the enhancement variance feature enhancement variance increasing rate (E3) demonstrated an obvious qualitative difference with respect to field strength. Size features, shape features besides irregularity, and morphology features did demonstrate some qualitative degree of agreement by field strength upon inspection, as expected.

There are some limitations to our study. For example, we limited our investigation to cancers of molecular subtype HR+/HER2−. We did this to reduce confounding variables, such as radiomic features that may vary between molecular subtypes of cancers [23,24]. Future investigations will study the robustness of radiomic features with respect to cancers of other molecular subtypes.

Secondly, the dataset of 612 unique lesions was comprised of lesions imaged at either field strength, which limited the first study to the use of population-based methods, as opposed to lesion-by-lesion comparisons.

Thirdly, our study focused only on investigating radiomic features in terms of their feature value distributions and individual feature classification performance, as extracted directly from clinically-acquired images. One limitation of this focus is that the extracted features are based on voxels within the segmented lesion volume. It would be useful to compare radiomics performance with respect to the impact of field inhomogeneity on image quality, which can differ by field strength [48-50] and correlate it to differences in features that depend upon signal and/or the signal to noise ratio [40,51,52]. Another limitation of this focus is that we did not investigate robustness of classification with respect to field strength when using machine learning methods that involve training classifiers with one set of lesions and testing them on another. This work will be the focus of a future study.

Fourthly, we did not investigate size features for potential robustness in the first study, because such a comparison can be conducted only on lesions imaged at both field strengths, a limitation of the large dataset which motivated our secondary demonstration study. The maximum lesion size of the lesions included in our smaller demonstration study differed in absolute terms somewhat between field strength (24% difference in lesion size for Subject 1, an average of 7% difference for the other six lesions). The difference in lesion size for Subject 1 may be due to the 90-day difference between image acquisitions, while the difference in size measurements for the other lesions is not unexpected. One previous prospective study of 20 patients diagnosed with ductal carcinoma in situ concluded that measurements of maximum lesion size as measured by radiologists demonstrated better correlation with pathology when the measurement was drawn from images acquired at 3.0 T versus 1.5 T [53]. It also suggested that all but one semi-quantitative kinetic curve feature failed to demonstrate significant difference across field strength, but in that study these features were calculated using only a 3×3×3 voxel volume for each lesion and thus did not capture the kinetics of the entire lesion. Given that the analysis here used automatically extracted radiomic features, including fully quantitative kinetic curve features drawn from the entire most-enhancing volume for each lesion, it is difficult to make comparisons between that study and the present one.

Additionally, our study was limited to investigation of a specific set of radiomic features extracted from DCE-MR images without the application of any harmonization step [54,55]. Our study design did not investigate differences in radiomic features themselves or compare the results against features extracted using software packages such as PyRadiomics [56]. In this work we aim to share which features may be robust across field strength at the point of image acquisition in the presence of clinically-relevant variations in scanner protocols, without data harmonization. Given the interest in the use of radiomic features extracted from other types of imaging sequences, such as diffusion weighted imaging [57-59] or ultrafast DCE-MR imaging [60], it could be additionally interesting to investigate the robustness of these radiomic features.

Finally, factors related to scanner manufacturer could be relevant in this study, but we were limited in the ability to control for them. Most lesions in this study were imaged using Philips Achieva or Philips Intera Achieva scanners, while the specific model name for the Philips scanner for two HR+/HER2− lesions imaged at 3.0 T was unknown. It has previously been demonstrated that Philips 3.0 T Ingenia scanner systems involve vendor-specific image scaling which can affect quantitative image analysis [61], and the authors of that work advised that image scaling should be considered for any images acquired using Philips scanners for quantitative analysis. This issue may extend to other Philips scanner systems, such as the ones used in this study. While the image scaling issue is not problematic for the analysis of DCE-MR images within a single series, which is how the images of individual lesions in this study were acquired, it could be the case that the image scaling introduced a variation in pixel scale between lesions and affected the comparison of feature values between the lesions. The image scaling may also vary for each acquisition, which could influence separate image acquisitions and resulting quantitative analysis of lesions which are then compared. At a more basic level, MR receiver gain is a fundamental factor in MR signal [62], as signal is proportional to scanner gain settings which may differ between imaging studies of different lesions. The effect is, of course, relevant to DCE-MR signal measurement [63]. It is possible that receiver gain settings were an additional factor that affected the assessment of the robustness of the radiomic features investigated here. The characterization of these factors was outside of the scope of these studies but will be of interest in future work.

However, despite these limitations, our reported results are useful because a variety of magnet strengths and protocols are used in clinical practice. Understanding which characteristics of radiomic features are impacted by field strength is relevant to the continued development of machine learning algorithms for computer aided diagnosis.

5. Conclusion

Potentially robust radiomic features have been identified for benign lesions and HR+/HER2− breast cancers imaged at two MRI field strengths, contributing to the development of radiomics in the context of different field strengths of image acquisition. Future studies will investigate robustness of radiomic features for benign lesions and HR+/HER2− breast cancers in the context of image quality measured from the full MR image, and for breast cancers of other molecular subtypes.

Figure 5:

Figures demonstrating similarity in feature values across magnet strength. Bland-Altman figures are shown for size, shape (except for irregularity), and morphology features, all of which by definition are expected to be independent of field strength, for each lesion imaged at two separate field strengths. Green circles indicate benign lesions; red circles indicate HR+/HER2− lesions. The numbers in the circles correspond to the subject numbers in Table 2. The data for maximum diameter (S4) is the same as given in Table 2.

Abbreviations

MR

magnetic resonance

DCE

dynamic contrast-enhanced

T1

the longitudinal relaxation time

HR

hormone receptor

HER2

human epidermal growth factor receptor 2

HIPAA/IRB

Health Insurance Portability and Accountability Act of 1996/ Institutional Review Board

FCM

fuzzy c-means

AUC

area under the (receiver operating characteristic) curve

ROC

receiver operating characteristic curve

BIRADS

Breast Imaging-Reporting and Data System

NMLE

non-mass-like enhancement

Appendix

Appendix A.

Radiomic features extracted from breast DCE-MRI images

Image Feature	Description
Size [11]
(S1) Volume (mm3)	Volume of lesion
(S2) Effective greatest dimension (mm)	Greatest dimension of a sphere with the same volume as the lesion
(S3) Surface area (mm2)	Lesion surface area
(S4) Maximum linear size (mm)	Maximum distance between any 2 voxels in the lesion
Shape [11]
(G1) Sphericity	Similarity of the lesion shape to a sphere
(G2) Irregularity	Deviation of the lesion surface from the surface of a sphere
(G3) Surface area/volume (1/mm)	Ratio of surface area to volume
Morphology [11]
(M1) Margin sharpness	Mean of the image gradient at the lesion margin
(M2) Variance of margin sharpness	Variance of the image gradient at the lesion margin
(M3) Variance of radial gradient histogram	Degree to which the enhancement structure extends in a radial pattern originating from the center of the lesion
Enhancement texture [10]
(T1) Contrast	Location image variations
(T2) Correlation	Image linearity
(T3) Difference entropy	Randomness of the difference of neighboring voxels’ gray-levels
(T4) Difference variance	Variations of difference of gray-levels between voxel-pairs
(T5) Angular second moment (energy)	Image homogeneity
(T6) Entropy	Randomness of the gray-levels
(T7) Inverse difference moment	Image homogeneity
(T8) Information measure of correlation 1	Nonlinear gray-level dependence
(T9) Information measure of correlation 2	Nonlinear gray-level dependence
(T10) Maximum correlation coefficient	Nonlinear gray-level dependence
(T11) Sum average	Overall brightness
(T12) Sum entropy	Randomness of the sum of gray-levels of neighboring voxels
(T13) Sum variance	Spread in the sum of the gray-levels of voxel-pairs distribution
(T14) Sum of squares (variance)	Spread in the gray-level distribution
Kinetic curve assessment [12]
(K1) Maximum enhancement	Maximum contrast enhancement
(K2) Time to peak (s)	Time at which the maximum enhancement occurs
(K3) Uptake rate (1/s)	Uptake speed of the contrast enhancement
(K4) Washout rate (1/s)	Washout speed of the contrast enhancement
(K5) Curve shape index	Difference between late and early enhancement
(K6) Enhancement at first postcontrast time point	Enhancement at first post-contrast time point
(K7) Signal enhancement ratio	Ratio of initial enhancement to overall enhancement
(K8) Volume of most enhancing voxels (mm3)	Volume of the most enhancing voxels
(K9) Total rate variation (1/s2)	How rapidly the contrast will enter and exit from the lesion
(K10) Normalized total rate variation (1/s2)	How rapidly the contrast will enter and exit from the lesion
Enhancement-variance kinetics [9]
(E1) Maximum variance of enhancement	Maximum spatial variance of contrast enhancement over time
(E2) Time to peak at maximum variance (s)	Time at which the maximum variance occurs
(E3) Enhancement variance increasing rate (1/s)	Rate of increase of the enhancement-variance during uptake
(E4) Enhancement variance decreasing rate (1/s)	Rate of decrease of the enhancement-variance during washout