The potential of predictive and prognostic breast MRI (P2-bMRI) – European Radiology Experimental

In contrast, semantic criteria are an integral part of routine breast MRI diagnostics [43]. This enables us to apply P2-bMRI in a large number of patients already today. In the following, we give an overview of how to apply semantic P2-bMRI criteria to clinical breast MRI protocols. Figures 1, 2, 3, 4, 5, 6, and 7 and Table 1 summarise key concepts of this approach.

Standardised reading setup for comprehensive diagnostic and predictive/prognostic breast magnetic resonance imaging (MRI) at one-stop shop. A female patient with suspicious amorphous segmental calcifications on the right breast at mammography, breast imaging reporting and data system (BI-RADS) 4 diagnostic category (not shown). MRI was performed also for preoperative staging due to suspicion of extended ductal carcinoma in situ. Diagnostic MRI shows an extensive heterogeneous segmental non-mass enhancement predominantly with plateau dynamic pattern. A noncircumscribed mass with washout and heterogeneous internal enhancement (BI-RADS 5 diagnostic category) is located centrally within the non-mass lesions. Relevant prognostic findings are here washout, skin thickening, invasion of the nipple, and diffuse ipsilateral oedema (see also Figs. 5 and 6). Semantic criteria correspond to the MRI phenotype of an aggressive invasive breast cancer. P2-MRI results were confirmed by postoperative pathological examination (invasive cancer NOS, G3, Ki-67+++, triple negative, node positive). Apparent diffusion coefficient map (A); unenhanced T1-weighted gradient echo with colour overlap of the dynamic curve on a pixel-by-pixel basis (green/yellow/red = persistent/plateau/washout) (B); first (C) and last (D) contrast-enhanced T1 GRE; diffusion-weighted imaging obtained with b = 800 s/mm² (E); T2-weighted turbo spin-echo (F); first (G) and last (H) contrast-enhanced subtractions. Diffusion-weighted imaging findings are highlighted in Fig. 7

Benefit of vascular analysis for predictive/prognostic breast magnetic resonance imaging. In A, B, and C, the enhancement patterns of three different breast cancers are colour coded on a pixel-by-pixel basis (green/yellow/red = persistent/plateau/washout). The maps reveal tumour heterogeneity, progressively increasing from A to C (no washout pixels in A, few of them in B, and many in C). Findings correspond to an increasing risk profile which was verified upon pathological and molecular analysis. Here, increasing vascularisation (CD 31 staining top row), cellular proliferation (Ki-67), and aggressiveness (grading) was demonstrated, and a less favourable receptor profile was evident from A to C. ER Oestrogen, her2neu Human epidermal growth factor receptor 2, PR Progesterone

Potential role of T2-weighed images: the lesion signal as a possible surrogate of water content. Signal intensity on T2-weighted images may reflect water content and warrants further scientific research. Compared to the surrounding parenchyma, a lesion is classified as hyperintense (A), isointense (B), or hypointense (C). In D, a mass lesion is identified on T2-weighted images (red arrow). T2-weighted signal intensity is assessed in comparison with the surrounding parenchyma. In D, just like the adjacent Cooper ligaments, the parenchyma displays less signal than the tumour. Correspondingly, the tumour is classified as “hyperintense”. Findings suggest the presence of an aggressive breast carcinoma phenotype. Predictive/prognostic findings were verified by immune histology revealing high-grade cancer with negative steroid receptors and elevated Ki-67 suggesting high cellular proliferation. Note perifocal and subcutaneous oedema (dotted arrows). In contrast, E displays a hypointense less aggressive carcinoma (red arrow, G2, positive steroid receptors, only Ki-67+)

The pivotal role of T2-weighed images: oedema. Any asymmetric ipsilateral T2-weighted signal increase not due to the tumour itself, cysts or artefacts is referred to as “oedema” in A, B, and C. Different oedema patterns are distinguished such as perifocal (A), full red arrow), subcutaneous (B, full red arrow), prepectoral (C, full red arrow), and diffuse (B, dotted red arrow). Oedema is considered among the best evaluated predictive/prognostic criteria and was associated with high grade and nodal-positive cancers as well as disease recurrence (see also Table 1)

Selected semantic parameters with known biological correlates. They can be readily implemented into clinical practice. Rim enhancement (A) was among the first morphologic parameters reported in the literature and reflects an aggressive cancers phenotype. This is also suggested by the adjacent vessel sign (A, magnification: dotted red arrow) [50]. Rim enhancement is thought to reflect central hypovascularity due to connective tissue, fibrosis, and/or necrosis. Necrosis sign (B) specifically depicts central colliquative (liquid) necrosis (B, magnification: full red arrow), characterised by a high signal intensity on T2-weighted images within the centre of the tumour. Invasion of the cancer into the nipple areolar complex is related to poor outcome. The semantic criterion described as “destruction of nipple line” (C) is best depicted on DCE images (C, magnification: red arrow). Further details including diagnostic performance of semantic parameters are provided in Table 1

Apparent diffusion coefficient (ADC) mapping (detailed analysis of the DWI already shown in Fig. 1). ADC values are extracted using quantitative region-of-interest-based measurements by standardised methods. ADC is given with (A) and without (B) colour overlap. ADC was applied to distinguish ductal carcinoma in situ (DCIS, orange) from invasive carcinoma (red). Note the presence of a benign lesion (a fibroadenoma), also correctly characterised by the ADC map (green), adjacent to the DCIS

Background parenchymal enhancement

The vascularisation of normal breast parenchyma is assessed by BPE [43]. In high-risk women, BPE has been identified as a prognostic imaging biomarker of breast cancer risk; women showing at least “mild” BPE are associated with significantly greater odds of future breast cancer (odds ratio: 2.1), which may be explained by deficient tissue repair mechanisms in this subgroup of women [46, 47] (Table 1). However, among average-risk women, the level of BPE is not associated with a higher risk of breast cancer [46].

The association of current BPE with breast cancer prognostic factors, such as higher mammographic density, steroid receptor status, and lymphovascular invasion has been reported in the literature [70]. Lim et al. [68] (hazard ratio 3.1) and Choi et al. [69] (postmenopausal; hazard ratio 3.9) independently reported the association of BPE with recurrence-free survival in average-risk patients with [70, 71]. These data emphasise the future potential of BPE as a genuine MRI imaging biomarker in the personalisation of breast cancer care.

Tumour enhancement: morphology and dynamics

A broad spectrum of semantic criteria is available to characterise breast tumour vascularisation. They may be applied to P2-bMRI as well [43, 50, 63, 72]. Neovascularisation is considered a key step in the process of carcinogenesis [73]. Patterns of neovascularisation can be assessed by microvessel density at traditional pathology examination, and this parameter is regarded a prognostic biomarker of breast cancer by itself [74]. Contrast enhancement is the basis for MRI diagnosis of breast cancer and is thought to reflect tissue vascularisation (Fig. 3) [2, 43]. Accordingly, many authors hypothesised that MRI enhancement patterns correlate with patient outcome and eventually may be used as imaging biomarkers [38, 40, 75].

Wash out is a key diagnostic criterion of the delayed enhancement phase but should also be approached as a prognostic biomarker [2, 43, 63]. It has been identified as a powerful tool to rule out the risk of metachronous metastasis (sensitivity and negative predictive value 100%; criterion, washout rate > 40%) [63]. Although these findings have to be validated in clinical trials, results highlight the potential of P2-bMRI parameters to estimate individual patient risk profile (Fig. 3, Table 1).

Breast cancer is a heterogeneous disease [37]. Volumetric analysis of MRI enhancement parameters investigate the composition of the entire tumour vasculature and are considered an imaging correlate of breast cancer heterogeneity [38, 76] (Fig. 3). Accordingly, the association of volumetric MRI parameters with histopathology and prognostic factors of breast cancer such as lymph node, hormonal receptor, and HER2 status has been demonstrated [40]. Building upon these results, correlation of volumetric MRI enhancement patterns with surrogate-free measure of patient outcome has been demonstrated in the meantime [38, 76]. For example, P2-bMRI has been shown to predict overall survival of breast cancer patients [76]. In a subsequent study, authors demonstrated that volumetric analysis of MRI enhancement yielded synergistic effects to conventional biomarkers [38]. Findings support the hypothesis that P2-bMRI can be used as add-on tool to further refine risk stratification of established prognostic biomarkers.

Rim enhancement is a classic diagnostic pattern of breast MRI [2, 43] (Fig. 6). Its prognostic value was early reported in the literature. Jinguiji et al. [61] investigated the relationship of this semantic MRI criterion with prognostic factors. Authors reported the significant association of rim enhancement with multiple prognostic factors such as lymph node metastasis, blood vessel invasion, steroid receptors, tumour size, and histological grade (G3 versus G1 or G2: diagnostic odds ratio 6.1; specificity 57.5%) (Table 1) [61]. Rim enhancement is thought to reflect central hypovascularity due to the presence of connective tissue, fibrosis, and/or necrosis in rapidly growing aggressive cancers [60, 61]. On the other hand, the necrosis sign is considered to indicate colliquative (fluid) necrosis, a pattern characterised by high signal intensity inside the cancer on T2-weighted scans [60]. As outlined in Table 1, necrosis sign has been described as one of the most specific semantic MRI criteria of high grade cancers (G3 versus G1 or G2 cancers: diagnostic odds ratio 3.7; specificity 94.3% (Table 1) [60].

Biomarkers from unenhanced breast MRI

T2-weighted signal intensity of a breast cancer is classified as hyper-, iso-, or hypointense compared to the surrounding breast tissue (Fig. 5, Table 1). In P2-bMRI, this semantic criterion serves as a predictor of tumour proliferation. Biologically less active desmoplastic tumours typically exhibit hypointense signal intensity on T2-weighted scans. In contrast, the presence of a T2-weighted hyperintense cancer suggests increased cellular proliferation and elevated Ki-67 expression (diagnostic odds ratio 2.2, specificity 59.8%) [62, 77].

While rim enhancement and signal intensity on T2-weighted scans investigate the gross anatomy of breast cancer, tumour microstructure can be investigated by diffusion-weighted imaging (DWI). DWI patterns are quantified by the apparent diffusion coefficient (ADC), which is typically used as a quantitative biomarker to aid differential diagnosis of suspicious breast lesions [5, 78, 79]. DWI could be used for P2-bMRI as well, and we regard the assessment of tumour invasiveness by ADC mapping as a promising clinical application [80]. Bickel et al. [80] reported higher ADC levels for ductal carcinoma in situ (DCIS) compared to invasive cancers. Personalisation of DCIS treatment is based on core biopsy samples, which are known to miss invasive tumour components in a relevant number of patients. Importantly, pure DCIS tumours require a different treatment strategy, which is why delaying correct diagnosis of invasive cancers should be avoided. To solve this dilemma, ADC may be used as a decision support tool [80]. Different to core biopsy, DWI examines the whole tumour, reducing the risk of sampling errors, which leads us to the following use case: if presurgical histology reveals DCIS, but ADC values (as well as patterns of contrast enhancement) are suggestive of invasive cancer, diagnosis of pure DCIS has to be questioned [80]. In this scenario, re-biopsy shall be considered and may avoid delayed diagnosis of invasive cancer as proposed by Bickel et al. [80] (Fig. 7).

Associated vascular findings

The adjacent vessel sign is a finding related to the macrovasculature of breast lesions. According to Dietzel et al. [70], the adjacent vessel sign reflects invasiveness of breast cancer. It indicates the presence of invasive cancer and is rarely seen in DCIS (diagnostic odds ratio 2.7, specificity 72.6%) [50] (Fig. 6, Table 1).

Whole breast vascular maps were investigated by Sardanelli et al. [72, 81] showing the association of an increased unilateral map with the presence of invasive cancers. This association was also shown to allow an increase in specificity using a 3-T magnet [82]. Martincich et al. [63] studied the variations of vascular maps in the context of primary systemic therapy. They showed that before therapy, vascular maps were asymmetrically increased ipsilaterally to the locally advanced breast cancer. After primary systemic therapy, vascular maps significantly changed only in the breast harbouring the cancer, with responders showing significantly more reduce vascular maps than nonresponders [63].

Associated nonvascular findings: oedema

The local tumour environment is recognised as a key factor in breast cancer development. It may be studied with semantic MRI criteria [83]. Oedema is defined as an associated finding in the BI-RADS lexicon, and it is characterised by T2-weighted signal increase within the local tumour environment [43, 54] (Fig. 5). It can be classified as perifocal, diffuse, subcutaneous, and prepectoral [39, 54,55,56, 84]. In general, the presence of oedema is indicative of aggressive cancer phenotypes, which is especially true for diffuse and prepectoral patterns [54, 55, 84]. According to Kaiser et al. [85], the latter is typically associated with lymph node metastases, lymphangitic carcinomatosis, and invasion of the chest wall. Subcutaneous oedema is specific of inflammatory breast cancer [84]. Whereas the majority of studies on the prognostic value of oedema used surrogates of patient outcome [54,55,56, 84], Cheon et al. investigated the impact of perifocal oedema on patient outcome [37]. Authors identified perifocal oedema as an independent biomarker of disease recurrence (hazard ratio 2.48) potentially improving the prognostication of disease recurrence by conventional biomarkers [39]. Meanwhile, the prognostic value of the semantic criterion “oedema” has been verified by numerous authors [86,87,88]. In particular, advanced techniques such as high-resolution diffusion-weighted imaging and radiomics showed promising potential to analyse peritumoural tissue, and they may further support the clinical impact ofP2-bMRI [89,90,91].

Tumour extent

Accurate assessment of anatomical tumour extent is the main rationale for preoperative breast MRI locoregional staging [10, 92]. However, this assessment provides significant prognostic information as well; tumour size is a key prognostic factor of breast cancer, and larger tumours are associated with a higher likelihood of worse outcome [92, 93]. Infiltration of breast cancer into associated structures such as the nipple areola complex or into the skin is associated with a poorer patient outcome [92, 94]. Accordingly, Dietzel et al. [52, 53] demonstrated that semantic criteria such as destruction of nipple line (diagnostic odds ratio 2.5, specificity 88.5%) or skin thickening (diagnostic odds ratio 5.9, specificity 94.5%) are associated with a poor prognostic profile such as that defined by the presence of locoregional lymph node metastases (Fig. 6, Table 1).

Occurrence of locoregional and distant metastasis deteriorates prognosis [93, 95]. Whole body MRI is an established tool to detect breast cancer metastasis, and its performance may be further improved by utilising the technology combining positron emission tomography and MRI [96,97,98]. While typical preoperative MRI aims to primarily assess ipsilateral tumour extension and the possibility of contralateral breast cancers, state-of-the-art scanner hardware could actually combine dedicated breast MRI examination with whole-body examinations as suggested by Kirchner et al. [98]. Since significant prognostic information can be derived from whole body MRI examinations, they are promising in the context for P2-bMRI as well [93, 95]. Different strategies exist for whole body examinations in breast cancer care. Kirchner et al. [98] proposed a complete whole body positron emission tomography/MRI staging. On the other hand, abbreviated protocols enabling screening for breast cancer metastasis are available as well [99, 100]. Requiring only 90 s of additional examination time, such protocols can be combined with a standard breast MRI within one single examination. Initial clinical data demonstrated promising results regarding both lymph node staging (positive predictive value of 100%, negative predictive value of 94.3%) and distant metastases screening (sensitivity 100%, specificity 98.3%) [99, 100].

Breast MRI phenotyping

Highly accurate breast MRI diagnosis is not achieved, until multiple parameters are assessed in concert [1, 2]. The same applies to P2-bMRI. Although individual parameters already enable prognostic assessment (Table 1), the relevance of P2-bMRI can be further specified when the lesion is analysed in the concert of multiple parameters [52, 99]. If these patterns correspond to a specific tumour biology, we refer to this feature combination as P2-bMRI phenotype. Accordingly, P2-bMRI phenotypes can provide actionable information, which is why we expect their key role in translating P2-bMRI into clinical practice.

An example of the use of P2-bMRI for phenotyping using artificial intelligence approach has been proposed by Dietzel et al. [101] (Fig. 8). Authors aimed to predict axillary lymph node metastases based on semantic MRI parameters of the index cancer. To support clinical application, prediction was based on a minimal number of MRI descriptors, and machine learning methods were used for this purpose [2, 99]. As expected, a single MRI parameter (skin thickening) was already able to predict the risk of axillary lymph node metastases (Fig. 8, Table 1). However, a reliable rule-out criterion (i.e., “no lymph node metastases”) could only be reached when three parameters were combined (Fig. 8, risk of nodal metastasis 0/56 = 0%).

P2-bMRI phenotypes are imaging patterns highly specific of a distinct tumour biology. They may be used as rule-in or rule-out criteria for clinical decision-making. Typically, P2-bMRI phenotypes are based on the assessment of multiple criteria in concert as in this example: here, a machine learning algorithm was used to identify phenotypes predictive of nodal-positive or nodal-negative stage (N+, N-). Semantic imaging criteria of the index lesion were used to predict nodal stage (for details, please see reference [42]). Classification results are presented as an intuitive and easy to follow decision tree. Accordingly, the “nodal-negative P2-bMRI phenotype” is characterised by a smooth lesion without oedema and without skin thickening. The positive likelihood of N+ is 0% for this P2-bMRI phenotype. Similar results can be achieved with other predictive/prognostic MRI methods, including artificial intelligence, each of them providing intrinsic advantages and disadvantages

The future

To translate P2-bMRI into P4 breast cancer care, three major challenges should be overcome.

First of all, methodological development of P2-bMRI needs to be refined. It may be achieved at the level of MRI data analysis. Radiomics of individual lesions (and machine learning applied to lesion radiomic data) as well as the use of convolutional neural networks applied to the whole image(s) may be regarded as most promising tools here [15, 19, 69, 102]. Yet, there is still considerable potential for improvement even only based on semantic criteria. Future development of P2-bMRI may also be achieved at the level of data acquisition, and magnetic resonance spectroscopy may be particularly promising here [65, 68].

However, methodological development by itself is not sufficient to translate P2-bMRI into clinical practice. Empiric evidence on P2-bMRI is generally derived from small, monocentric, and retrospective studies. Clinical application will request validation of P2-bMRI in a real-world oncological setting before adopting the methods. This calls for dedicated interdisciplinary, large, multicentre studies, perhaps also randomised controlled trials.

Finally, the breast imaging community itself should try to be a strong promoter in the process of translating P2-bMRI into clinical practice. The prognostic potential of imaging biomarkers is not yet sufficiently recognised outside the field of radiology. Only when radiologists and nuclear medicine physicians dedicated to breast imaging will be successful in convincing key stakeholder such as patients, clinical colleagues, healthcare providers, and MRI vendors, P2-bMRI will start to be an integral part of routine breast cancer care.