Background
Breast imaging settings and potential impact of data-driven tools
Imaging plays a decisive role in the detection, management, and follow-up of breast
disease. By definition, breast imaging is multi-modal, meaning that a combination
of different imaging methods is employed as suited to the clinical workflow. For a
better understanding of the results and decisions that may be impacted by data-driven
tools, a structured description is helpful.
Screening
Despite ongoing controversies around screening tests in general and breast cancer
screening specifically, most expert societies recommend secondary breast cancer prevention
by screening mammography for the early detection of breast cancer [1]. Mass screening applies the same test (mammography) to the eligible population,
that is, with some variations, women aged 50–70 years. In recent years, medicine has
moved to a more personalized approach that recognizes that diagnostic and therapeutic
interventions have an individual impact [2]. This also holds true for breast cancer screening: women differ regarding their
risk profile and screening mammography is less effective in dense breasts as compared
to non-dense breasts as lesions are more likely to be masked by breast parenchyma.
In addition, breast density is one independent risk factor that is easily accessible
by mammography [3]. A woman with dense breasts is thus at double risk: higher risk of developing cancer
compared to the average population and risk of missed cancer by mammography. On the
other hand, a subgroup of women may exhibit such a low risk of breast cancer that
screening is so unlikely to have an impact on morbidity and mortality that mammography
screening may not be required. Thus, one future approach to solve this issue is to
use data-driven tools including automated image analyses to stratify women into those
who require screening mammography, those who require supplemental or alternative screening
tools such as breast MRI, and those who may not require screening at all ([Fig. 1]). The mass screening setting is characterized by a low prevalence of breast cancer
and thus an overwhelming proportion of negative (= healthy) cases [4]. Image data-driven tools could help to identify cases that are definitely benign
and thus reduce the workload and consequently possible reader mistakes as a result
of performing a tiring and monotonous task [5]. The latter could be avoided by computational image-data analysis pointing out evident
pathology. A positive finding is a screening recall, usually confirmed or refuted
by a second reader and then sent to an assessment center ([Fig. 1]). Again, data-driven tools can help in this step by acting as a second reader, providing
an objective risk of cancer estimation, reducing unnecessary recalls ([Fig. 1]).
Fig. 1 Arrows represent decisions, dashed arrows logical steps. Possible interaction steps
of data-driven decision support tools are highlighted as grey A.I. arrows.
Abb. 1 Pfeile stehen für Entscheidungen, gestrichelte Pfeile für logische Schritte. Mögliche
Interaktionsschritte von datengesteuerten Werkzeugen zur Entscheidungsfindung sind
als graue A.I.-Pfeile hervorgehoben.
Assessment
Both screening recalls in asymptomatic women as well as symptomatic women are assessed
by further imaging and, if required, minimally invasive image-guided biopsies. The
physician has to decide whether the recall or symptomatic finding requires further
imaging or biopsy (and the method of image guidance) or can be downgraded and what
follow-up tests and intervals are necessary. Both the choice of further imaging as
well as the decision provide ample opportunity to employ data-driven decision-support
tools ([Fig. 1]). From a value-based health care perspective, avoidable biopsies and avoidable additional
tests provide value for the patient while at the same time saving resources.
Therapy guidance and monitoring
Breast cancer treatment is driven by molecular subtypes that are usually determined
using immunohistochemically examined tissue specimens from image-guided biopsies.
A significant proportion of breast cancer cases undergo neoadjuvant therapy before
surgery that can even be omitted or at least delayed in the case of multimorbid patients.
As neoadjuvant treatment is a main cost driver in breast cancer diagnosis and treatment
[6], imaging tests could play a pivotal role in adapting treatment [7]. While therapy could be deescalated in early complete responders, therapy regimes
could be changed in non-responders. While there is currently no solid evidence regarding
which technology and imaging markers should be used when, it is conceivable that the
complexity of the matter would, again, profit from data-driven decision-support tools.
Key buzz words
Before we address the topic in a narrative overview on the current literature, three
important terms are introduced and briefly discussed from the breast imaging perspective.
Breast imaging biomarkers
The term “biomarker” was initially restricted to biologic molecules [8]. The much broader current definition considers any “defined characteristic that
is measured as an indicator of normal biological processes, pathogenic processes,
or responses to an exposure or intervention, including therapeutic interventions”
as a potential “biomarker” [9]. This conceptual change opened the door to the evaluation of radiologic examinations
as an “imaging biomarker” [10].
In clinical practice and breast cancer-related research, biomarkers are used as surrogate
outcome measures [11]. Surrogate parameters are of great importance, because direct endpoints (e. g. overall
survival) are very difficult to capture in clinical trials [12]. Typical applications of biomarkers are prediction and monitoring of therapy response (e. g., ki67 determination during ongoing neoadjuvant chemotherapy)
as well as prognosis of patient outcome (risk stratification using genetic microarrays) [13]
[14]
[15]
[16]
[17]. Accordingly, biomarkers are considered pivotal in precision medicine [18].
Imaging biomarkers may be classified as quantitative, ordinal, and qualitative. Breast cancer size remains one of the most important surrogates of patient outcome
and treatment response, which is why radiological measurement of tumor dimensions
is a classic quantitative breast imaging biomarker [19]
[20]. Beside morphologic evaluations, breast imaging offers numerous advanced methods
of functional tissue assessment. For instance, the tissue microstructure may be quantitatively
investigated by the apparent diffusion coefficient (ADC) [21]
[22]. Distinguishing invasive and in situ breast cancer growth is pivotal both for therapy
planning and for the estimation of patient outcome. According to Bickel et al., the
ADC may be used as a quantitative breast imaging biomarker of breast cancer invasiveness
[22].
Breast density has been established as one of the most important independent risk
factors of breast cancer [3]
[23]
[24]. In the clinical routine breast density is visually assessed on a four-level scale
and thus may be considered an ordinal imaging biomarker [24]. In addition, a variety of methods have been described to assess breast density
ranging from fully automated to quantitative techniques [25]
[26]
[27]
[28].
Qualitative breast imaging descriptors are typically used to establish the diagnosis of breast cancer [24]. However, many of these patterns could be associated as well with the underlying
tumor biology and may be considered qualitative biomarkers. For instance, perifocal
edema is predictive of malignancy but also suggestive of unfavorable prognosis such
as the presence lymph nodes metastasis [29]
[30]
[31].
Imaging Biomarker Panels
The discriminatory power of individual breast imaging biomarkers can be further improved
if they are combined with one another into imaging biomarker panels. Various methods
have been investigated for this purpose, with artificial intelligence and radiomics
being most promising [14]
[15]
[32]
[33]
[34]. For instance, the occurrence of lymph node metastasis may be further improved by
the joint analysis of qualitative and quantitative breast imaging biomarkers applying
artificial neural networks or radiomics [33]
[35]
[36]
[37].
Radiomics
Among the methods for combined analysis of imaging biomarker panels, radiomics have
arguably gained the greatest attention [38]. Since the introduction of the term, to date a total of 3145 articles on radiomics
have already been published, of which the majority (49.4 %, 1555/3145) are related
to breast imaging [39]. This illustrates both the huge interest among the scientific community in radiomics-related
research and the dominant role of breast imaging.
Radiomics approaches breast images not as mere pictures but as a source of comprehensive
data containing viable patient information not perceivable with our human eyes. The
analytic radiomic workflow starts with segmentation of the tissue of interest, followed by extraction of a large number of tissue characteristics expressed as arithmetic functions, of
which the most important ones are chosen during statistical feature selection and applied for machine-learning-based model building. The output of this workflow is an algorithm or model that can diagnose or predict
the intended target condition or property. The result of the radiomic analysis or
the specific radiomics signature is typically expressed as a numeric value (the predicted
probability of the target condition being met) and may be regarded as a quantitative
imaging biomarker [32].
In clinical practice radiomics may be used for virtually any use case such as distinguishing malignant from benign breast lesions, the prediction of patient outcome, tumor biology, and response to neoadjuvant chemotherapy and examples from the literature are highlighted in [Fig. 2], [3]
[36]
[40]
[41]
[42]
[43].
Fig. 2 Application of radiomics and AI to improve the diagnostic accuracy of breast imaging.
Stelzer et al. examined the radiomic signatures of microcalcifications in 235 patients
[69]. All findings were assessed as suspicious in conventional visual analysis (BI-RADS
IV) and therefore bioptically confirmed. A basic radiomic workflow was applied starting
with marking of a region (dashed lines) of interest for data extraction, selection
of radiomic features, building and testing the radiomics model. As a result, the authors
received a specific radiomic signature, based upon which a biopsy could be safely
omitted in up to 45.7 % of the patients. Shown are BI-RADS IV findings correctly classified
as benign (A: benign papilloma without atypia; B: fibrocystic changes) or malignant based on the radiomic signature (C: invasive ductal cancer, D: ductal carcinoma in situ). Results are available in real time, not requiring additional
imaging or invasive procedures. Therefore, radiomic signatures may be considered accurate,
workflow-friendly, and cost effective which is why they provide value to patient care
(modified and reprinted with permission [69]).
Abb. 2 Stelzer et al. untersuchten die radiomischen Signaturen von Mikroverkalkungen in
235 Patienten [69]. Alle Befunde wurden in der konventionellen visuellen Analyse (BI-RADS IV) als verdächtig
bewertet und daher bioptisch bestätigt. Es wurde ein Basis-Workflow angewandt, der
mit der Markierung einer für die Datenextraktion interessanten Region (gestrichelte
Linien), der Auswahl der radiomischen Merkmale, dem Aufbau und der Prüfung des radiomischen
Modells begann. Als Ergebnis erhielten die Autoren eine spezifische radiomische Signatur,
auf deren Grundlage bei bis zu 45,7 % der Patienten eine Biopsie sicher vermieden
werden konnte. Dargestellt sind BI-RADS-IV-Befunde, die anhand der radiomischen Signatur
korrekt als gutartig (A: gutartiges Papillom ohne Atypie; B: fibrozystische Veränderungen) oder bösartig klassifiziert wurden (C: invasiv duktales Karzinom, D: duktales Karzinom in situ). Die Ergebnisse liegen in Echtzeit vor und erfordern
keine zusätzliche Bildgebung oder invasive Verfahren. Daher können radiomische Signaturen
als genau, Workflow-freundlich und kosteneffektiv angesehen werden, weshalb sie für
die Patientenversorgung relevant sind (modifiziert und mit Genehmigung nachgedruckt
[69]).
Fig. 3 a Application of radiomics to improve risk stratification of breast cancer patients.
Extracts of the radiomics workflow-based breast MRI data are displayed above. Volumetric
analysis of dynamic enhancement parameters was performed providing surrogates of tumor
heterogeneity and cancer vascularity [17]. Such MRI derived parameters are closely linked to pathophysiology, facilitating
interpretation of the radiomic model, a step not always feasible with classic texture
parameters. The resulting radiomic signature was integrated into an established prognostic
model (Nottingham prognostic index: NPI) yielding NPI+. Upon benchmarking analysis
of NPI versus NPI+, the additive benefit of the radiomic signature is evident. Corresponding
Kaplan Meier curves illustrate a better risk stratification of NPI+ compared to standard
NPI (modified and reprinted with permission [17]). b Application of radiomics to improve risk stratification of breast cancer patients
(benchmarking analysis).
Abb. 3 a Oben werden Auszüge aus den auf dem Radiomics Workflow basierenden MRI-Daten der
Brust dargestellt. Es wurde eine volumetrische Analyse dynamischer Anreicherungsparameter
durchgeführt, die Surrogate der Tumorheterogenität und der Krebsvaskularität liefert
[17]. Solche Parameter sind eng mit der Pathophysiologie verbunden und erleichtern die
Interpretation des radiomischen Modells; ein Schritt, der mit klassischen Texturparametern
nicht immer möglich ist. Die resultierende radiomische Signatur wurde in ein etabliertes
Prognosemodell (Nottingham Prognostic Index, NPI) integriert und ergab NPI+. Nach
Benchmarking-Analyse von NPI versus NPI+ wird der additive Nutzen der radiomischen
Signatur deutlich. Die Kaplan-Meier-Kurven veranschaulichen dort eine bessere Risikostratifizierung
von NPI+ im Vergleich zu Standard-NPI (modifiziert und nachgedruckt mit Genehmigung
[17]). b Anwendung von Radiomics zur Verbesserung der Risikostratifizierung von Brustkrebspatientinnen
(Benchmarking Analyse).
Artificial intelligence
Next to radiomics, artificial intelligence (AI) is the emerging topic in radiology
[44]
[45]
[46]
[47]
[48]. Both methods aim at processing information otherwise hidden from human interpretation
and may be used in combination [44]
[49]. Just like radiomics, AI can be used for virtually any clinical use ranging from
distinguishing malignant from benign breast lesions to the prediction of direct patient
outcome [44]
[50]
[51]. Yet, unlike radiomics, the definition of AI is rather vague. This might be due
to the fact that the term (non-artificial) “intelligence” already lacks a generally
accepted definition [52]. So commonly, AI is used as an umbrella term, summarizing various disciplines of
computer science, all of which aim to mimic certain cognitive functions and can be
as simple as basic multivariable classification algorithms [44]
[45]
[46]
[47].
In breast imaging the most investigated subdiscipline of supervised AI is machine
learning [44]. Without being explicitly programmed, machine learning algorithms aim to detect
(“learn”) patterns in imaging data and to link these imaging characteristics to tissue
characteristics [44]
[46]. A large variety of methods belong to the machine learning family, such as artificial
neural networks (ANNs), support vector machines (SVMs), but also classic statistical
procedures, such as decision trees and regression analysis [46]. Each method provides inherent advantages and has been successfully applied to breast
imaging research [14]
[33]
[53]
[54].
Without any doubt, AI in breast imaging goes far beyond supervised machine learning.
Further pivotal AI methods such as unsupervised learning and deep learning have already
shown high promise for the future development of our field [44]
[49]
[55]
[56]
[57]. The complexity and importance of this topic call for a dedicated publication summarizing
the state of the art of AI in breast imaging [18]
[44]
[51].
Current status and future development: A modality-based approach
Mammography
The idea of using computer algorithms to process mammography-derived imaging data
is not as new as it seems. In fact, this research topic has been investigated since
the 1980s [58]. The first studies on computer-aided detection (CAD) systems aimed at supporting
the radiologist in the early detection of breast cancer on digital mammography. However,
the widespread use of these systems was initially limited by their low specificity
and high number of false-positive findings, which reduced the positive effect on the
screening readings [59]. The introduction of advanced AI methods and the availability of large databases
together with the framework of international multireader studies reshaped and improved
the performance of CAD systems [60].
Lesion characterization
Recent studies showed that AI-based systems may have a performance not inferior to
that of experienced breast radiologists [60]
[61]. Instead of completely replacing the human reader, AI might also be used to switch from double reading to single reading in screening. In this setting the second reader will be substituted by AI [50]. Another promising clinical use case of AI is to triage screening mammograms as cancer-free. In this scenario, AI might potentially reduce
the radiologist’s workload and gain extra time for the read of more cases and/or the
assessment of challenging cases [5]
[62]
[63]. Furthermore, the same algorithms applied for mammography seem to perform equally
well for digital breast tomosynthesis (DBT). This finding might be essential for reducing
the increased workload related to the use of DBT for instance in screening [64]
[65].
Irrespective of these encouraging results, a long list of methodological, empirical,
ethical, and forensic issues needs to be solved [66]
[67]. Accordingly, we do not expect these systems to completely replace radiologists
in the foreseeable future. Moreover, the most realistic use in the diagnostic setting
is the application of radiomics and AI as decision support tools in the assessment
for instance in challenging cases, such as the assessment of microcalcifications and
BI-RADS IV lesions as demonstrated in [Fig. 2]
[68]
[69].
Beyond lesion characterization
Increased breast density is a factor that increases the risk of missing a cancer on
mammography, due to tissue superimposition, and also is among the most important risk
factor for developing breast cancer [3]. Historically breast density has been evaluated in a semi-quantitative manner using
the American College of Radiology Breast Imaging Reporting and Data System [24]. Despite the overall good inter-reader agreement in the assessment of breast density
with this semiquantitative method [25], a more accurate and repeatable evaluation of density might improve and standardize
the identification of women with dense breast tissue and a subsequently increased
risk of breast cancer [26]. Several machine learning-based methods have been developed in the last years, and
several studies have shown an association between breast cancer risk and breast density
measured using these automated devices [27]. These automated calculations are still subject to variation related to image acquisition
(i. e., compression), vendor, and other physical properties [28], but provide a reliable tool for a standardized and repeatable evaluation of breast
density [26]
[27].
Beyond the assessment of breast density, radiomics and AI may also predict lymph node
metastasis and HER2neu status [35]
[36]
[40]
[42]. Notably, the additional information derived from the application of an intravenous
contrast agent seems to be able to further improve the performance of these algorithms
warranting future investigation [36]
[40].
Ultrasound
Breast ultrasound (US) is a valuable adjunct to mammography, with a high sensitivity
for breast cancer [70]. With the use of advanced techniques, besides B-mode, a US examination can provide
information about further tissue properties, such as stiffness (elastography) or vascularity
(Doppler, contrast-enhanced US (CEUS)). Radiomics features can be extracted from all
these techniques, not only in order to differentiate benign from malignant lesions,
but also to identify factors with a prognostic and/or predictive value for breast
cancer.
Lesion characterization
Attempts to use breast US for the extraction of radiomic data begun in the early ‘90 s
with the performance of texture analysis studies and yielded good results for the
distinction of benign from malignant breast lesions. Already in 1993, Garra et al.
achieved a sensitivity of 100 % and a specificity of 73 % by using textural features
derived from co-occurrence matrices [71]. Initial attempts aimed at the development of CAD systems, which at that time demonstrated
an increase in sensitivity at the cost of a decreased specificity, especially for
experienced readers [72].
Advances in artificial intelligence methods during the last decade allowed for more
sophisticated use of US-based radiomic data. Marcon et al. applied a machine learning
algorithm to textural features acquired from automated breast ultrasound and achieved
an accuracy of 90.7 % for the differentiation of benign from malignant lesions [73]. Lee et al. developed a radiomics score for the discrimination of fibroadenomas
from triple-negative carcinomas, which demonstrated an AUC of up to 0.853, noting
however that the performance of the score was dependent on the type of US machine
used [41].
Several attempts have been made to apply radiomics to advanced US techniques in order
to improve lesion characterization. Zhang Q. et al. used radiomics features derived
from strain elastography examinations and coined the term “sonoelastomics” [74]. In their study, seven such features achieved an AUC of 0.917 for the identification
of breast malignancies. A more recent study by Zhang X. et al. compared deep-learning
based radiomics scores acquired from B-mode US and shear-wave elastography (SWE) with
the results of the BI-RADS assessment and quantitative SWE parameters and found a
significant increase in the diagnostic performance by using radiomics scores, reaching
an AUC of 1 for both [75]. On the preclinical level, Theek et al. performed a radiomics analysis of CEUS data
from different mouse tumor models, achieving correct classifications in 82.1 % of
cases [76].
Irrespective of these encouraging results and similar to mammography, ultrasound-derived
radiomics is unlikely to replace the human reader in the near future and the biggest
clinical advantage is to be expected with respect to decision support [77].
Beyond lesion characterization
Several attempts have been made to identify breast cancer-related imaging biomarkers
based on US-derived radiomics features. Guo et al. reported that different molecular
subtypes of breast cancer show differences in several radiomics features [78]. In their study, triple-negative carcinomas could be distinguished from other subtypes
with an AUC of 0.760, based on a feature combination. On the other hand, a series
of studies have focused on the prediction of axillary lymph node metastases in breast
cancer patients [49]
[79]. Yu et al. developed a nomogram combining US radiomics and clinical features, which
could accurately predict the lymph node status with an AUC of 0.81 [80], while Zheng et al. was able to accurately distinguish between a high and a low
axillary burden using a combination of clinicopathologic data and radiomics features
derived from B-mode US and SWE [55].
Challenges and perspectives of US-based radiomics
Although US is routinely implemented in breast diagnostics, its use for radiomics
studies clearly lags behind that of other modalities, such as MRI or mammography [38]. One of the main reasons is the inherent high inter-examiner variability of US images.
US image quality depends on both the acquisition process as well as the examiner’s
experience while there is also substantial variability between different US devices,
as has also been demonstrated by different studies. Moreover, the quality of US images
is usually further reduced due to the presence of noise and artifacts. Finally, there
is a paucity of large publicly available, labeled, anonymized datasets that would
be mandatory for radiomics studies. Keeping these limitations in mind, breast US with
its ubiquitous presence, its practicality, and the possibility to acquire a multitude
of different morphological and functional information at a low cost has significant
potential for application in the definition of breast tumors radiomic signatures and
may thus play a substantial role in precision medicine.
PET
Positron emission tomography (PET)-CT with fluorodeoxyglucose (18F-FDG) is mostly
indicated for inconclusive CT findings and high-risk patients and plays only an ancillary
role in the staging of breast cancer [81]. Yet, after the introduction of PET-MRI, the interest in PET for local and distant
breast cancer staging has increased.
Local Staging
The use of 18F-FDG PET-MRI has been proposed for local staging in women with breast
cancer, but the additional information derived from the PET examination was only marginal
in the first studies [82]
[83]. Further analysis showed that MRI and PET-derived data could yield relevant additional
information. In particular, several studies applied radiomics to the PET-MRI images
of the breast and proved that AI could be used to characterize breast lesions and
define tumor aggressiveness [84]. Similar to breast density in mammography, background parenchymal enhancement and
background parenchymal uptake may be regarded as potential imaging biomarkers for
breast cancer incidence [85]. Texture analysis and radiomics could also play a role in the evaluation of tumor
aggressiveness as well in the prediction of response to neoadjuvant chemotherapy [86]
[87]
[88]. However, only few single center studies are available on PET MRI-derived imaging
biomarkers for local breast cancer staging. More research in this field is needed.
Distant Staging
PET-MRI is currently rarely used for whole-body staging of newly diagnosed breast
cancer. Increasing evidence suggests that the examination could improve sensitivity,
as compared to PET-CT, but the limited availability and the high costs of the examination
suggest that careful patient selection is warranted [89]. Currently, no large studies have been performed using AI to detect or characterize
distant metastasis in patients with breast cancer, but animal studies suggest that
the combined information from MRI, PET and AI could improve the early detection of
metastasis [90].
MRI
Lesion characterization
In 2003, Gibbs et al. reported the potential of quantitative multivariate texture
analysis to distinguish benign from malignant lesions in breast MRI [91]. This paper anticipated the key concepts of lesion characterization by breast MRI
radiomics by many years and key results have been validated ever since [32]
[54]
[92]
[93]
[94]
[95].
For instance, machine learning was used to generate the Kaiser score. Based on qualitative
BI-RADS MRI descriptors, the Kaiser score is a typical ordinal imaging biomarker.
It allows estimation of the individual likelihood of breast cancer and specific underlying
tumor biology [54]
[96]. Results have been validated by multiple centers in different clinical scenarios,
underlining the potential of AI-derived decision support in clinical patient management
(for instance [94]).
Whereas the Kaiser score is based on visually extracted qualitative features, numerous
authors investigated quantitative features for breast MRI lesion characterization
[92]
[93]
[95]. Yielding similar results compared to the Kaiser score, Parekh and Jacobs developed
a radiomic framework using advanced machine learning and multiparametric MRI. In a
cohort of 124 patients, they were able to semiautomatically distinguish benign from
cancerous lesions achieving an AUC≤ 0.91 [95]. As small lesions in breast MRI can be difficult to characterize by visual analysis
[97], Gibbs et al. investigated radiomics signatures to characterize subcentimeter breast
lesions. Applying support vector machine analysis, authors reported AUC levels ranging
from 0.75–0.81 [93].
Benchmarking analysis refers to the comparison of radiomics and AI results with an
established gold standard (see [Fig. 3]) [98]
[99]
[100]. Unlike many researchers (e. g. [93]
[95]), Bickelhaupt et al. included a benchmarking analysis in a pivotal paper on radiomics
signatures derived from diffusion MRI. The authors report promising performance of
the radiomics models (AUC ≤ 0.85), which were, however, clearly outperformed by human
readers (AUC = 0.96) [92]. Such benchmarking analysis enables us to estimate the practical value of radiomics
and AI in the clinic by providing context. Consequently, benchmarking is considered
mandatory for clinical validation of novel imaging biomarkers [98]
[99]
[100].
Beyond lesion characterization
Breast MRI seems ideally suited for imaging biomarker research, as it provides comprehensive
multiparametric data on pathophysiological tumor characteristics [15]
[34]
[101]
[102]. Consequently, a large number of papers are available, demonstrating the predictive
and prognostic potential of radiomics and/or AI-derived MRI analysis [9]
[101].
Prediction of treatment response is pivotal in the management of breast cancer and radiomics. Granzier et al. summarized
the literature on breast cancer response prediction to neoadjuvant systemic therapy
using MRI-based radiomics. The authors report a high diagnostic accuracy for this
diagnostic task (AUC ≤ 0.94) [15]. The prediction of therapy response also depends on the receptor status of the breast
cancer [103]. In this respect, the results of Wu et al. are remarkable. They were able to predict
the molecular subtype based on BI-RADS features in a multi-modal radiomics analysis (f1-score < 87.9 %)
[101]. Based on steroid receptor status, HER2neu status, and ki67 analysis, the molecular
subtype has significant prognostic and predictive power and significantly influences
the choice of therapy for breast cancer, which is why the results of Wu et al. are
clinically relevant [101]
[103].
One of the most important prognostic factors for patient outcome is lymph node status
[19]. Based on visually extracted MRI descriptors, artificial intelligence may be able
to identify MRI patterns suggestive of lymph node metastases [33]
[37]. Continuing this early research, Liu et al. investigated quantitative radiomics
for the prediction of sentinel lymph node metastasis. The diagnostic accuracy exceeded
the level achievable with qualitative features (AUC = 0.74 vs. 0.81) [33]
[102]. Notably Liu et al. could further improve the radiomics model by integrating clinicopathological
biomarkers (AUC = 0.81 vs. 0.87) [102].
Multigene assays represent advanced prognostic and predictive biomarkers of breast cancer. A disadvantage
of this approach are the high costs, which is why Bhargava et al. propose alternative
methods to safely forgo multigene assays [104]. According to the results of Li et al., MRI radiomics correlate to a certain extent
with commercially available multigene assays (r = 0.5–0.56) [105]. In analogy to the approach of Bhargava et al., MRI radiomics could also be used
to select patients that would benefit from an advanced multi-gene assay [104]
[105].
Already in 2003, Boné et al. reported that the combined analysis of qualitative enhancement
pattern can predict disease-free survival of breast cancer
[106]. Results have since been reproduced by many authors such as Kim et al., Pickles
et al., and Baltzer et al. using various radiomics methods and study endpoints (disease-free,
recurrence-free survival, and overall survival) [17]
[43]
[107]
[108]. Some authors benchmarked MRI-derived imaging biomarkers with established prognosis
tools [17]
[106]. By benchmarking MRI-derived radiomics with the Nottingham prognostic index (NPI),
we could demonstrate that our prognostic radiomics model performed equally well in
predicting overall survival of breast cancer patients. More important, the combined
analysis of MRI radiomics and the NPI improved the predictive accuracy. These findings
underline the potential of creating synergistic effects, if biomarkers are used in
combination [17].
Discussion and conclusions
Considering radiological examinations not as mere images but as a source of data has
become the key paradigm in the diagnostic imaging field. This change of perspective
allows radiologists to apply methods derived from computer science, to realize innovative
clinical applications, and to refine already established methods. In this context,
the terms “imaging biomarker”, “radiomics”, and “artificial intelligence” are of pivotal
importance. These methods promise noninvasive, low-cost (e. g. in comparison to multigene
arrays) and workflow-friendly (automated, only one examination, instantaneous results)
delivery of clinically relevant information [38]. For all areas of breast imaging, exemplary studies were highlighted reporting a
variety of clinical use cases. Ultimately, these data may provide value by optimizing
patient management by guiding individualized breast cancer diagnosis and treatment
in the age of precision medicine.
Yet, a closer look into the literature shows that there are expectations regarding
promising technologies that are unlikely to be fulfilled in the foreseeable future
[109]
[110]. In order to close the remaining research gaps, considerable efforts from the scientific
community are necessary. However, the successful translation of this new imaging paradigm
into clinical action requires much more than empirical evidence.
Many stakeholders are making claims on the field such as mass media, patients, and
colleagues from non-radiological disciplines and the industry. Taking the lead in
this innovation process goes far beyond the role of individual research institutions.
It requires communication strategies, joint research projects between academia and
vendors, and educational formats accessible to large radiological societies. We are
convinced that only in this way can radiomics and AI be translated into the clinical
routine and develop their promising advantage in order to improve patient care. In
breast cancer diagnosis and treatment, we have highlighted typical workflow steps
that would benefit from these new tools, providing value by improved selection of
diagnostic tools and therapeutic strategies.