Differentiation of invasive ductal and lobular carcinoma of the breast using MRI radiomic features: a pilot study

Introduction

Breast cancer (BC) is the most frequently diagnosed cancer among women worldwide. With an expected 2.3 million new cases, or 11.7% of all cancer cases worldwide in 2020, lung cancer has surpassed lung cancer as the most common cause of cancer incidence.¹ According to epidemiological studies, by 2030, there will likely be roughly 2 million BC patients worldwide, according to epidemiological studies.² Early and precise identification and characterization of cancers are crucial because of their incidence and therapeutic significance.

Mammography and Ultrasonography are frequently used for breast lesion detection, screening and diagnostic purposes.^3,4 Breast Magnetic Resonance Imaging (MRI) is performed regularly to better detect primary and recurrent tumors, characterize them, and assess the patient’s response to therapy. Dynamic contrast-enhanced (DCE) MRI is an important component of the MRI-Breast protocol, which involves serial capture of strong T1 weighted images during intravenous administration of contrast. It provides information about tumor vascularity, and enhancement curves are frequently used to increase specificity (greater than 90%) for diagnosing cancer.^5–8 Accurately diagnosing invasive ductal carcinoma (IDC) and invasive lobular carcinoma (ILC) presents significant challenges, primarily due to overlapping clinical, radiological, and histological features. Diagnostic doubts are exacerbated by mammographic ambiguities, equivocal biopsy results, and challenges in differentiating between the two subtypes. The distinctive marker for ILC, E-cadherin, can be stained with immunohistochemically, but the results can be ambiguous. Additionally, the genetic and molecular variability within each subtype makes classification even more difficult. Furthermore, the whole extent of the tumor may not always be captured by the diagnostic techniques used today, such as mammography, ultrasound, and biopsy, which could result in sample errors and incomplete assessments. These drawbacks highlight the necessity of continued research, thorough multidisciplinary approaches, and diagnostic technology developments in order to improve patient outcomes and increase the precision of IDC and ILC diagnoses.^9–12

Radiomics is a popular area of study in the processing and analysis of medical images. Radiomics provides more information than the visual and qualitative patterns that radiologists can see with their unaided eyes by extracting a large number of quantitative imaging features from medical images. On routine imaging examinations performed in cancer patients, radiomic feature (RF) characteristics can non-invasively evaluate intratumoral variability.^13,14 The use of RF in patients with BC is a novel and developing area of translational research. RF in BC has been extensively employed in research settings with the hope that it may eventually improve diagnosis and characterization.^15,16

There are significant methodological variations in research involving RF and artificial intelligence (AI) techniques, such as machine learning (ML) based on radiomics, and there is potential for methodological advancement and standardization to improve study quality in BC. There is a lack of reproducibility and validation in radiomic studies.¹⁷ Our literature review revealed a few studies that used RF from the DCE sequence of breast MRI for differentiating malignant lesions such as invasive ductal carcinoma (IDC) and invasive lobular carcinoma (ILC). Hence, our study investigated whether RF obtained from DCE-MRI would aid in the differentiation of IDC and ILC.

Methods

First, this was a retrospective study. The study was commenced upon approval from the Institutional Ethical Committee of Kasturba Medical College and Hospital, Manipal, India on 6^th August 2022 (IEC 202/2022). Informed consent was waived since the data was collected from a publicly available database.

Image segmentation and RF extraction

The DICOM MRI images of the Axial T1 DCE were uploaded into 3D slicer (version 4.10.2) and the regions of tumor were manually (slice by-slice) delineated by radiologist who had experience more than 10 years (Figure 1). The RF was extracted from the axial T1 dynamic contrast-enhanced sequence. ROIs were defined across the entire tumor using DCE-MRI images with the strongest enhancement phase. The observer specifically chose phase (3.92 ± 1.22-on average), to segment the analyzed lesions, out of the six or seven phases offered by the Axial T1 Dynamic contrast enhanced sequences, across different patients, where the breast mass was more visible than in the backdrop. The radiologist was blinded to histopathological reports.

Figure 1. Example of the segmentation of the lesion in axial T1 (3D) dynamic contrast enhanced sequence.

Results

In our study, we analyzed the RF extracted from DCE-MRI. Our study extracted 107 features from the DCE-MRI sequence for each subject. The mRMR technique identified 15 RF features that were relevant for differentiating between IDC and ILC Table 3. Of the 15 selected features, 10 DCE MRI-based RF were significant for differentiating between IDC and ILC. A total of 58 cases were included in the study. The mean age (years) of the patients with IDC and ILC was 48 ± 11.15 and 59 ± 11 years, respectively.

MRI radiomic features

For 3D DCE MRI sequence, Gray level dependence matrix (GLDM) gray level variance for IDC and ILC was 288.6 ± 136.4 and 1187.0 ± 342.5 (p < 0.001), Gray level co-occurrence matrix (GLCM) square for IDC and ILC was 118.5 ± 63.15 and 492.4 ± 277.6 (p < 0.001), GLCM difference average for IDC and ILC was 4.472 ± 2.694 and 20.45 ± 10.11 (p < 0.001), GLCM cluster tendency for IDC and ILC was 270.3 ± 81.47 and 1170.8 ± 145.5 (p < 0.001), GLCM interquartile range for IDC and ILC was 286.2 ± 52.8 and 1630.8 ± 332.9 (p < 0.001), first order robust-mean absolute deviation for IDC and ILC was 117.9 ± 18.2 and 403.5 ± 179.0 (p < 0.001), first order entropy for IDC and ILC was 4.776 ± 1.656 and 6.063 ± 1.697 (p < 0.001), first order variance for IDC and ILC was 44324.8 ± 3799.0 and 438604.1 ± 39979.8 (p < 0.001), Gray level run length matrix (GLRLM) gray level variance for IDC and ILC was 35.64 ± 31.20 and 741.40 ± 173.3 (p < 0.001) and GLRLM run entropy for IDC and ILC was 5.496 ± 1.677 and 6.814 ± 1.65 (p < 0.001) (Table 4 and Figure 2).

Table 4. DCE-MRI based radiomic features for differentiating the malignant lesions of the breast.

MRI radiomic features	IDC (n = 30)	ILC (n = 28)	p-value
	(Mean ± SD)
DCE GLDM Gray Level Variance	288.6 ± 136.4	1187.0 ± 742.5	<0.001
DCE GLCM Sum Squares	118.5 ± 63.15	492.4 ± 277.6	<0.001
DCE GLCM Difference Average	4.472 ± 2.694	20.45 ± 10.11	<0.001
DCE GLCM Cluster Tendency	270.3 ± 81.47	1170.8 ± 745.5	<0.001
DCE GLCM Interquartile Range	286.2 ± 152.8	1630.8 ± 732.9	<0.001
DCE First Order RMAD	117.9 ± 18.2	403.5 ± 279.0	<0.001
DCE First Order Entropy	4.776 ± 1.656	6.063 ± 1.697	<0.001
DCE First Order Variance	44324.8 ± 3799.0	438604.1 ± 39979.8	<0.001
DCE GLRLM Gray Level Variance	35.64 ± 31.20	741.40 ± 473.3	<0.001
DCE GLRLM Run Entropy	5.496 ± 1.677	6.814 ± 1.65	<0.001

Figure 2. MRI-based significant radiomic features for differentiating IDC and ILC from axial T1 (3D) dynamic contrast enhanced sequence.

(a) GLDM Gray Level Variance; (b) GLCM Sum Squares; (c) GLCM Difference Average; (d) GLCM Cluster Tendency; (e) GLCM Interquartile Range; (f) First Order RMAD; (g) First Order Entropy; (h) First Order Variance; (i) GLRLM Gray Level Variance; (j) GLRLM Run Entropy.

Accuracy measures of radiomic features

The accuracy measures of RF for differentiating between IDC and ILC are listed in Table 5. For 3D DCE MRI sequence, GLRLM gray level variance at cut off value of 42, had higher sensitivity (SN 97.21%), specificity (SP 96.2%), and area under curve (AUC 0.998; GLCM interquartile range at cut off value of 357 had higher SN (95.24%), SP (97.31%), and AUC of 0.968; and GLCM difference average at a cut off value of 6.5, had higher SN 95.72%, SP 96.34% and AUC of 0.983 compared to GLDM gray level variance (SN 91.02%, SP 89.72%), GLCM sum squares (SN 85.91%, SP 82.72%), first order variance (SN 81.54%, SP 82.72%), GLCM cluster tendency (SN 81.77%, SP 80.13%), first order RMAD (SN 80.12%, SP 79.23%), first order entropy (SN 75.24%, SP 72.23%), GLRLM run entropy (SN 75.47%, SP 77.83%) (Figure 3).

Table 5. Area under curve, sensitivity, specificity, and cut-off value for MRI based Radiomic features for differentiating malignant lesions of breast (IDC and ILC).

	Radiomic Features	Area under the curve	Sensitivity (%)	Specificity (%)	Cut off value
DCE MRI	DCE GLDM Gray Level Variance	0.887	91.02	89.72	408.5
	DCE GLCM Sum Squares	0.877	85.91	82.72	169.5
	DCE GLCM Difference Average	0.983	95.72	96.34	6.5
	DCE GLCM Cluster Tendency	0.792	81.77	80.13	334
	DCE GLCM Interquartile Range	0.968	95.24	97.31	357
	DCE First Order RMAD	0.768	80.12	79.23	216
	DCE First Order Entropy	0.709	75.24	72.23	5.5
	DCE First Order Variance	0.805	81.54	82.72	780637
	DCE GLRLM Gray Level Variance	0.998	97.21	96.2	42
	DCE GLRLM Run Entropy	0.698	75.47	77.83	5.508

Figure 3. Receiver Operating Characteristic (ROC) curves for differentiating IDC and ILC from Axial T1 (3D) Dynamic contrast enhanced sequence.

(a) GLDM Gray Level Variance; (b) GLCM Sum Squares; (c) GLCM Difference Average; (d) GLCM Cluster Tendency; (e) GLCM Interquartile Range; (f) First Order RMAD; (g) First Order Entropy; (h) First Order Variance; (i) GLRLM Gray Level Variance; (j) GLRLM Run Entropy.

Discussion

In the current study, we explored the usefulness of DCE-based RF compared to MRI for differentiating malignant lesions of the breast, such as IDC and ILC. Breast MRI is superior to mammography and ultrasonography for early detection of BC. IDC and ILC are the most common subtypes of malignant cancer. Differentiation of ILC from IDC is quite difficult, as there are very few differences between them. The morphological and dynamic contrast kinetic characteristics of IDC and ILC did not differ considerably from each other. The correct identification of ductal and lobular carcinoma helps in the improved management and overall survival analysis of the patient.^22,23

Previous studies have focused on differentiating benign and malignant lesions of the breast using radiomic models in MRI.^24–26 However, few studies have been conducted to differentiate invasive carcinomas of the breast, such as IDC and ILC, using RF from dynamic contrast sequences. The integration of radiomic models in normal radiological practice will be extremely beneficial for non-invasive diagnosis and clinical management of invasive BC in the future.

Our study found ten RF useful in differentiating IDC and ILC. We noticed that categories such as GLDM, GLCM, first order, and GLRLM showed significant differences in differentiating ductal and lobular carcinoma of the breast. Of these categories, GLRLM (Gray level variance AUC 0.998) and GLCM difference average) and interquartile range features were the best predictors for differentiation between IDC and ILC. GLCM features consider how pixels are arranged in space. These features describe the texture of a picture by calculating the frequency of pixel pairings with distinct values and a specific spatial relationship. GLRLM features are used to describe the length of successive pixels with the same grey level value in terms of pixels. GLCM and GLRLM are important markers for assessing tumor heterogeneity and for better characterization of malignant subtypes.

Waugh et al.²⁷ in their study noticed that all co-occurrence RF had higher accuracy (71.4% and AUC –0.745) in differentiating ductal and lobular carcinoma than entropy features (64.7% and AUROC –0.632). Holli et al.²⁸ noticed that the co-occurrence RF of subtraction first images was more statistically significant than that of other features (Table 6). They also achieved a classification accuracy of 100% using first subtraction and contrast series using nonlinear and linear discriminant analyses. Our study also noticed that co-occurrence matrix features, such as sum squares, difference average, and cluster tendency, exhibited good accuracy in differentiating ductal and lobular carcinomas. In addition to co-occurrence matrix features, we also noticed additional categories showing statistically significant differences, which were not reported by Holli et al.²⁸ and Waugh et al.²⁷

A study by Fusco et al.²⁴ observed that kurtosis and skewness (AUC = 0.71) in X-ray mammography and range, energy, entropy, and gray-level non-uniformity (GLN) of the GLRLM from DCE-MRI were the best predictors for differentiating benign and malignant lesions. Niu et al.²⁵ they studied the accuracy of RF extracted from digital mammography (DM), digital breast tomosynthesis (DBT), diffusion (DWI), and DCE MRI for the characterization of breast lesions and noticed that the RF extracted from DWI and DCE MRI yielded higher AUC, SN, and SP with DCE having the upper hand compared to DWI, and lower AUC, SN, and SP were noted from DM. Militello et al.²⁶ reported that shape-based features such as least axis length, flatness, and elongation, GLCM-based features such as joint energy, GLRLM features such as short run emphasis, and Gray level size zone (GLSZM) of size-non-uniformity from DCE-MRI exhibited the highest SN and SP in characterizing breast lesions.

A study by Lafci et al.²⁹ noticed that Gray level zone length matrix (GLZLM) features had the highest accuracy (AUC = 0.718) in distinguishing Luminal A and B types of ductal carcinoma. They also noted that Luminal B tumors had a larger volume than luminal A tumors as they are aggressive and require intense chemotherapy, and observed shape-based features such as voxel volume showed significant differences between A and B. In our study, we also noted a larger voxel volume for ILC (5144 ± 306.5) than for IDC (3899 ± 684.3); however, we did not notice a statistically significant difference because of the smaller sample size. Literature suggests that ILC has a larger volume and is more aggressive compared to IDC.³⁰

Our study has the following limitations. First, we did not utilize machine learning and deep learning classifiers for the prediction of ductal and lobular carcinomas due to the small sample size. Second, longitudinal studies with large sample sizes and machine learning methods were used to further validate the results. Third, as the RF is obtained using manual segmentation, it is time-consuming and subjective.

The limited sample size of a pilot study may introduce biases into the study, which examined the use of MRI radiomic characteristics to distinguish between invasive ductal and lobular cancer of the breast. This would restrict the applicability of the findings. There is a chance that overfitting and inflated diagnostic accuracy will occur, so interpretation should be done carefully. Variations in imaging protocols, equipment, and radiomic feature extraction techniques between different institutions can lead to problems with reproducibility. Larger, multicenter studies should be prioritized in future research initiatives to improve generalizability and sample variety. The study's dependability will be increased by addressing potential biases through stringent statistical validation, using standardized imaging techniques, and using external validation datasets. Additionally, examining the influence of biological and molecular factors on radiomic features could lead to a more thorough comprehension of the distinction between invasive ductal and lobular carcinoma based on imaging, which could improve the accuracy and clinical applicability of breast cancer diagnosis.

Conclusions

Classification of BC into histological subgroups is a dynamic process. Our study suggested an RF-based method for ductal and lobular carcinoma characterization using T1 dynamic contrast sequence, which might give radiologists additional value for decision making in a noninvasive method and could be utilized clinically for malignant BC classification. Due to potential biases and limited application, the pilot study's small sample size investigating MRI radiomic characteristics for differentiating between IDC and ILC cancer may be problematic. Cautious interpretation is necessary because of potential problems including overfitting and overstated diagnostic accuracy. Larger, multicenter studies are crucial for better generalizability since variations in imaging methods throughout institutions may present reproducibility issues. Future studies should focus on external validation datasets, standardized imaging methods, and rigorous statistical validation to improve reliability. They should also look into the influence of biological components to gain a more thorough understanding of breast cancer diagnosis.