Skip to main content
NCBI home page
NIHPA Author Manuscripts logoLink to NIHPA Author Manuscripts
. Author manuscript; available in PMC: 2024 Mar 18.
Published in final edited form as: PET Clin. 2023 Jun 17;18(4):567–575. doi: 10.1016/j.cpet.2023.05.002

AI-Enhanced PET and MR Imaging for patients with breast cancer

Valeria Romeo 1, Linda Moy 2, Katja Pinker 3
PMCID: PMC10947491  NIHMSID: NIHMS1973723  PMID: 37336693

Abstract

New challenges are currently faced by clinical and surgical oncologists in the management of breast cancer patients, mainly related to the need for molecular and prognostic data. Recent technological advances in the field of diagnostic imaging and informatics have led to the introduction of functional imaging modalities, such as hybrid PET/MRI, and artificial intelligence (AI) software, aimed at the extraction of quantitative radiomics data, which may reflect tumor biology and behavior. In this chapter, the most recent applications of radiomics and AI to PET/MRI are described to address the new needs of clinical and surgical oncology, represented by non-invasive tumor profiling, lymph node status and risk recurrence assessment, and early prediction of the response to primary systemic therapy.

Introduction

Breast cancer is the most frequent solid tumor affecting women worldwide [1]. Notably, while the five-year-survival rate for patients diagnosed with stage I breast cancer approaches 100%, patients with later-stage breast cancer often have a poor prognosis [2]. Hence, much effort has been expended to develop advanced strategies not only for treatment but also early diagnosis. Some of the exciting developments in the field of breast cancer in the recent decades stem from the increased understanding of tumor biology and the emergence of targeted treatment approaches. In addition, there have been many technological developments in the field of diagnostic imaging allowing physiological data to be obtained beyond morphological data [3].

Such developments in diagnostic imaging involve techniques such as positron emission tomography (PET) and advanced magnetic resonance imaging (MRI) sequences including diffusion-weighted imaging (DWI) which interrogates the microstructure, perfusion-weighted imaging (PWI) which interrogates neoangiogenesis, and magnetic resonance spectroscopy (MRS) which interrogates cancer-related metabolites [4]. Although most of these techniques are still confined to the research realm, they are promising to assess different aspects of tumor biology and several quantitative parameters derived from these techniques have been linked with tumor aggressiveness and metastatic potential. In this context, simultaneous PET/MRI, the newest hybrid imaging modality, has shown great promise, especially for local/distant staging and treatment monitoring [5–7]. In addition, the ability of PET/MRI to obtain biologically related quantitative data can be further enhanced with artificial intelligence (AI) [8], employing either traditional handcrafted radiomics coupled with machine learning (ML) or deep learning (DL). The underlying principle of AI applications in medical imaging is that imaging features can be extracted from medical images that encode both simple patterns and many higher-order patterns not discernable with the naked eye and that can be useful for diagnosis, prognostication, and prediction [8–12].

In this chapter, we explain how the most advanced imaging modalities employed in breast cancer patients, PET and MRI, can be empowered by radiomics and AI to meet the new needs of clinical and surgical oncology in breast cancer management. In the first section, we present a summary of the current trends of breast cancer management and the questions to be answered. In the second section, we present a state-of-art review on radiomics and AI applications to PET and MRI, addressing the questions raised in the first section.

Clinical needs for breast cancer

Tissue characterization and molecular profiling

Current knowledge shows us that breast cancer is a heterogeneous disease, related to the identification of breast cancer molecular profiles with different degrees of biological aggressiveness and molecular targets, leading to different patient outcomes and prognoses [13,14]. As such, each patient represents a specific case, requiring a personalized therapeutic approach, and the preoperative knowledge of the breast cancer molecular profile is essential to establish the most appropriate treatment. Such assessment is currently performed through the analysis of a histological sample of tumor tissue obtained by a core biopsy, a procedure not free from risk that, additionally, does not provide information on the whole lesion [15]. Depending on the breast cancer molecular profile, different therapeutic approaches can be offered, with primary systemic therapy (PST) being increasingly used. In addition, the shifting of breast cancer to a different molecular subtype can occur based on tumor clonal diversity and the development of treatment resistant clones, such that the non-invasive assessment of tumor heterogeneity through a “virtual biopsy” would be extremely advantageous.

Preoperative assessment of axillary lymph node involvement

While the available imaging modalities have reached a high accuracy for breast cancer diagnosis, thanks especially to the high sensitivity of dynamic contrast-enhanced (DCE) MRI, the preoperative assessment of axillary lymph node involvement remains an unsolved issue. Indeed, for the latter, there is high variability in diagnostic performance between the available imaging modalities, with ultrasound still representing the most sensitive modality (87%), even if affected by a variable specificity (53%–97%) [16]. At present, sentinel lymph node biopsy (SLNB) remains the gold standard for assessing axillary lymph node involvement [17]. This issue is critical as a re-assessment of treatment strategy is required when axillary lymph node involvement unrecognized at imaging is found after SLNB, especially for patients who could have benefited from PST. On the other side, a more conservative axillary surgical approach is preferred in selected cases, in the context of de-escalation treatment, to avoid side effects related to axillary node dissection, such as arm lymphedema [18,19]. According to current guidelines, axillary node dissection can be avoided not only in patients with negative lymph node after SLNB, but also after PST in clinically node-positive patients presenting as cN0 after treatment and in patients with fewer than three positive axillary lymph nodes at SLNB who are receiving breast/axillary radiation [18,20].

Primary systemic therapy

Recurrence risk assessment and early prediction of response

In locally advanced breast cancer, PST can make surgery feasible in non-operable cases as well as allow a more conservative surgical approach. Furthermore, with PST, the in vivo assessment of response is still possible, allowing the possibility to change/stop treatment in non-responder cases, with a tangible impact on patient outcome and prognosis [21]. As such, even considering the new trends on de-escalation surgery [22], PST is still increasingly recommended even in operable, but selected cases [23]. Indeed, PST-related issues must be considered, such as systemic/cardiac toxicity, psychological implications, and the possible occurrence of tumor progression related to chemotherapy-induced changes of the breast cancer molecular profile [24]. Therefore, a careful assessment of PST clinical indications is mandatory. So far, PST is highly recommended in specific breast cancer subtypes such as HER2+ breast cancer, in combination with targeted drugs, and triple-negative breast cancer, for which no targeted approaches are currently available [23]. For luminal subtypes, hormone treatment is indicated in luminal A breast cancer, while the usefulness of PST in luminal B breast cancer is still debated. At present, patients with luminal B breast cancer are referred to PST in locally advanced, non-operable cases (e.g., with axillary lymph node involvement or T4 stages) or to multi-gene tests (e.g., Oncotype DX, Mammaprint, PAM50) on tumor sample after surgical excision to assess the risk of tumor recurrence and therefore the cost/effectiveness of an adjuvant systemic treatment [22, 23]. Similarly, since functional response to treatment seems to precede morphological changes, the possibility for early prediction of the response to PST based on multiparametric/multimodal quantitative imaging obtained preoperatively or during early imaging assessment through AI-enhanced analysis could further stratify and select patients for whom the cost-benefit ratio is advantageous. A summary of clinical and surgical needs is illustrated in Figure 1.

Figure 1.

Figure 1.

Open in a new tab

Clinical areas of interest for AI applications in breast cancer, represented by tumor characterization/molecular profiling, which, along with the preoperative assessment of axillary lymph node involvement and multigene tests, helps in defining clinical indications for neoadjuvant chemotherapy. An accurate definition of axillary status, in terms of the number of involved lymph nodes, and neoadjuvant chemotherapy allow for a more conservative surgical approach, the so called “de-escalation” treatment. One of the most promising and fascinating clinical applications of PET/MRI functional imaging coupled with AI is also the early prediction of the response to neoadjuvant chemotherapy, which would help in further selecting the ideal candidates among patients with breast cancer.

Clinical needs addressed from the “AI perspective”

AI basics concepts: machine and deep learning

The rationale behind AI is that, if properly instructed, computers can learn to analyze a multitude of data to make predictions and improve their performance with the experience. Two different approaches can be used for this purpose: traditional radiomics and ML, and DL.

Traditional radiomics and ML extracts quantitative imaging features that are used to identify a phenotypical fingerprint or “radiomics signature.” The cancer is annotated by expert readers or automated software reflecting the distribution of pixels at different complexity levels. The radiomics pipeline for ML studies is usually made of different steps, including image pre-processing, segmentation, radiomics features extraction and selection, and running of the ML algorithms [27].

DL employs a complex network inspired by the human brain architecture to devise its own features. Currently, in medical image analysis, DL algorithms use convolutional neural networks (CNN), which comprise multiple layers of processing designed to optimize millions of variables, the so-called weights and biases, to extract hierarchical patterns, to retain the most important information and use them for classification [28,29]. The majority of DL models utilize a supervised learning approach in which training is done using a multitude of labeled examples which can be on different levels (exam, breast, pixel). While big datasets are not necessarily required by ML systems, they are essential for DL studies, which must learn features from the data. Consequently, high computational time and costs are required for running DL software, depending on the architecture and the size of the dataset. An advantage of DL over ML is that it can process a huge amount of data, but is considered a “black box.” Once a model is developed, an essential condition is that results obtained in a “training” population have to be validated preferentially in an external “test” set, i.e., from a different institution. Simplified ML and DL pipelines are illustrated in Figure 2.

Figure 2.

Figure 2.

Open in a new tab

Illustration of machine learning and deep learning radiomics pipelines.

Tissue characterization and molecular profiling

Although breast MRI has a high sensitivity for breast cancer detection, the possibility to use AI to non-invasively discriminate benign from malignant breast lesions, and in addition to determine the molecular profile of breast cancer, is attractive and has been recently explored using different imaging modalities. In highly suspicious breast lesions (e.g., BI-RADS 4 and 5), a combined MRI and PET approach could be of value to provide tumor diagnosis, profiling, and staging at the same time. Therefore, initial experience was recently published on the use of ML for breast cancer diagnosis and phenotyping using simultaneously acquired PET and MRI images.

In 2018, an unsupervised clustering based on PET and MRI radiomics features was performed by Huang et al., extracting a total of 84 radiomics features of 113 patients. Three groups were identified, significantly associated with tumor grade, stage, breast cancer subtypes, and disease recurrence status. This preliminary experience suggested that both MRI and PET could be able to decipher breast cancer biological behavior, while providing imaging biomarkers predictive of tumor recurrence [30]. In a further investigation, an ML-based model for breast cancer diagnosis using a combination of radiomic features; quantitative MRI diffusion and perfusion parameters; and PET parameters was proposed. The integrated model combining mean transit time and mean apparent diffusion coefficient (ADCmean) with radiomic features extracted from PET and ADC images obtained an area under the curve (AUC) of 0.983 [31].

The same group also developed a method for breast cancer phenotyping to discriminate triple negative from other breast cancer subtypes. After the calculation of quantitative parameters and radiomic features from PET and MR images (Figure 3), different combinations of such data were explored. The best performing ML method employed radiomic features extracted from ADC and PET images and obtained an AUC, sensitivity, and specificity of 0.887, 79.7%, and 86%, respectively [32].

Figure 3.

Figure 3.

Open in a new tab

Examples of 2D ROI placement for the extraction of quantitative parameters (mean transit time; plasma flow; volume distribution; ADC mean; and SUVmax, mean, and minimum) (A–C), and whole tumor segmentation for radiomics features (first, second, and higher order) extraction (D–G) from primary breast cancer tumor lesions on DCE (A,E), DWI (B,F), PET (C,G), and T2‐weighted (D) images. Reprinted under a CC BY 4.0 license from: Romeo V, Kapetas P, Clauser P, Baltzer PAT, Rasul S, Gibbs P, Hacker M, Woitek R, Pinker K, Helbich TH. A Simultaneous multiparametric 18F-FDG PET/MRI Radiomics Model for the Diagnosis of Triple Negative Breast Cancer. Cancers (Basel). 2022 Aug 16;14(16):3944. doi: 10.3390/cancers14163944. PMID: 36010936; PMCID: PMC9406327.

Similarly, Umutlu et al. developed different ML models for breast cancer characterization, in terms of molecular subtype (luminal A vs luminal B, luminal A vs others), ER/PgR status, HER2 expression, Ki67 levels, and tumor grade, using combination of MRI and/or PET features. As a result, the AUC of the developed models ranged from 0.771 (tumor grade prediction model, based on PET features) to 0.97 (Ki67 levels prediction model, based on MRI and PET features) [33].

Overall, the combined use of MRI and PET could provide morphological tumor information along with functional and histological data, thus representing a “one-stop-shop” tool for a comprehensive breast cancer diagnosis and staging. Such an approach would have a significant implication for patients’ management. Indeed, the performance of a “virtual biopsy” would dramatically reduce patients’ discomfort as well as allow a comprehensive assessment of whole-lesion heterogeneity and the monitoring of tumor feature changes during PST, to assess for tumor resistance and progression.

Preoperative assessment of axillary lymph node involvement

At present, little evidence is available on the use of ML/DL applied to PET and MRI for the preoperative evaluation of axillary lymph node status. The rationale of current investigations is that axillary dissemination of breast cancer cells strongly depends on primary lesion heterogeneity, such that the majority of AI analyses are conducted to extract radiomic features from primary breast cancer lesions. In this task, preliminary investigations report a good performance of ML algorithms using radiomic feature extracted from breast cancer lesions on MRI and PET images in the prediction of axillary status (positive vs negative), with an AUC, sensitivity, and specificity of 0.810, 63.8%, and 82.2%, respectively [33].

Besides preliminary studies exploring the potential of both MRI and PET for the assessment of axillary lymph node status, several investigations have been published over the past years to assess the individual contribution of these modalities. Regarding MRI, a recent systematic review and meta-analysis by Zhang et al., including 13 studies on the use of ML applied to DCE MRI images for the prediction of axillary lymph node metastasis in 1618 patients, reported a pooled sensitivity, specificity, diagnostic odds ratio, and AUC of 0.82, 0.83, 21.56, and 0.89, respectively [34]. Studies on the use of ML applied to PET have also been recently conducted. In an interesting investigation, Chen et al. aimed to develop and validate a ML model applied to 18F-FDG PET/CT for the prediction of clinically occult axillary lymph node metastasis (cN0) in 180 patients [35]. Different ML algorithms were employed, with random forest resulting as the best performing one, showing a mean AUC of 0.817, and a mean accuracy of 81.2%.

The possibility to combine PET with additional clinical and imaging data has also been explored by Cheng et al., who developed different ML model, including clinical (physical examination), histological (ER status), imaging (ultrasound and PET) and radiomics data [36]. In the validation set, the combined model, made of six clinicopathologic factors and five radiomic features extracted from dedicated PET, yielded the highest diagnostic performance (AUC = 0.93, sensitivity = 92.11, specificity = 83.67, accuracy = 87.36). ML has also been recently employed to combine clinically assessable features on MRI and PET and possibly compare/overcome the performance of expert radiologists [37]. A total of 303 PET/MRI examinations were collected from three different institutions. The authors found no difference between the performance of ML and radiologists in assessing the presence of axillary lymph node metastases, with a diagnostic accuracy of 91.2% and 89.3%, respectively. When using MRI alone, the accuracy was 87.5% for both ML and radiologists, showing no significant difference with that of PET/MRI. Among PET/MRI features, the most relevant were FDG uptake and lymph node size. With an adjusted threshold, a decision tree was built which could help in reducing the number of invasive procedures, such as sentinel lymph node biopsy, in 68.2% of cases.

With all these premises, it seems that information provided by histology, clinical examination, conventional and radiomic image features of different imaging modalities could allow an accurate, non-invasive detection of axillary lymph node metastasis in breast cancer. On this basis, ML is promising as an effective tool for the elaboration of these complex data, supporting clinicians in their clinical practice as a potential clinical decision-making instrument. Since surgical treatment changes according to the number of affected axillary lymph nodes, a significant improvement could be to preoperatively predict the involvement of more than two axillary lymph nodes, in a way that patients could be directly addressed to axillary lymph node dissection.

Primary systemic therapy

Recurrence risk assessment

New genomic tests are currently available to assess the recurrence risk in patients with breast cancer and therefore to identify those who would benefit from systemic treatment, especially in hormone receptor-positive subtypes. While these tests are currently performed on post-operative surgical specimens, the prediction of recurrence score from pretreatment imaging examinations would be advantageous, allowing high-risk patients to undergo the systemic treatment in a neoadjuvant setting, with all the related, previously discussed benefits. Among these tests, Oncotype DX score is one of the most widely used. Indeed, several investigations have been carried out for its early prediction, using semantic MRI features and multivariate models [38,39]. ML techniques have been recently employed for this purpose along with clinical variables and multiparametric radiomics, obtaining an AUC of 0.89 in discriminating between low and intermediate/high-risk groups [40]. In a recently published paper, ML was applied to MRI alone for the non-invasive prediction of Oncotype DX score [41]. Despite the use of a limited sample size (248 patients from a publicly available dataset), encouraging findings are reported (accuracy in the test set= 63 and AUC of 0.66), suggesting a possible role of this modality for recurrence prediction. DL systems were also used for discriminating between patients at low and intermediate/high risk of breast cancer recurrence, with an overall accuracy of 84% [42]. Still, more evidence is needed to fully explore the potential of AI as a clinical decision-making tool for breast cancer recurrence risk assessment.

Early prediction of the response to PST

The possibility for early prediction of response to PST using cytotoxic chemotherapy in breast cancer has extensively been investigated. In this task, radiomics is one of the most promising tools, thanks to its ability to describe imaging heterogeneity patterns invisible to human assessment, and before any morphological changes can be appreciated. Among the available imaging modalities, MRI and PET are the best candidates, because both exams capture functional, quantitative data related to tumor neoangiogenesis (DCE-MRI), cellularity (DWI) and metabolism (PET).

Specifically, neoadjuvant chemotherapy acts at the level of tumor cell density, due to its cytotoxic effect that increases the extravascular/extracellular space and that reduces tumor vascularization and permeability. In addition, the anti-angiogenetic effect of neoadjuvant chemotherapy affects tumor metabolism, which also tends to decrease early in patients who respond to PST. As a result, several studies have assessed the usefulness of quantitative parameters reflecting such functional changes as well as radiomic features for the detection of early PST-related changes. Consequently, systematic reviews and meta-analyses are now available to summarize the current evidence on this matter.

In a recent paper by O’Donnell et al., different breast MRI radiomics methods were analyzed and compared using a network meta-analysis, including quantitative functional parameters, radiomic features, and different time points of MRI examinations (before, during, and after PST) [43]. The authors demonstrated that radiomic features performed better than quantitative parameters, and during- and post-PST were the best time points for the prediction of the response. Another systematic review and meta-analysis, which included 34 studies on MRI radiomics, showed a pooled AUC of 0.78 (95% CI: 0.74–0.81) for the prediction of pathological complete response to PST. However, the authors also found substantial heterogeneity in the included studies [44].

The role of 18F-FDG PET/CT for the prediction of response to PST has also been widely explored. In a recent systematic review, the promise of PET-derived radiomic features was reported, with possible improvements when clinical features were also included in the model [45]. In a retrospective study by Li et al., 100 18F-FDG PET/CT examinations were collected and employed for the extraction of 2210 PET-derived radiomics features [46]. The random forest ML classifier was employed and obtained a prediction accuracy of 0.857 (AUC = 0.844) on the training set and 0.767 (AUC = 0.722) on the test set. When patient age was also included in the model, the accuracy of the predictive model increased to 0.857 (AUC = 0.958) and 0.8 (AUC = 0.73) in the training and test set, respectively.

More recently, studies combining clinical imaging data and/or radiomic features obtained from both PET and MRI images were conducted, using either ML or DL approaches. In a first experience, Choi et al. investigated the ability of DL applied to 18F-FDG PET/CT and MR images for the prediction of pathological complete response after PST in 56 patients, comparing its performance with that of conventional quantitative data obtained from pre-treatment and interim (after the first PST cycle) PET and DWI examinations (Figure 4) [47]. Among the quantitative data, early variation in the standard uptake value (SUV) showed the highest AUC of 0.805 (95% CI: 0.677–0.899). Of note, the accuracy of PET data improved after the application of DL (from 0.687 to 0.980), but this did not occur for DWI data. A first experience of ML applied to pretreatment simultaneous 18F-FDG PET/MRI for the prediction of the response has been recently reported by Umutlu et al. A total of 73 18F-FDG PET/MRI examinations were retrospectively collected and analyzed for the extraction of 101 radiomics features [48]. The support vector machine algorithm was used, with a 5-fold cross validation, obtaining the highest accuracy, sensitivity, and specificity (0.8, 81%, and 73.8%, respectively) when radiomic features from both 18F-FDG PET and MRI were combined. In a subgroup analysis according to the different breast cancer molecular subtypes, the best performance (AUC = 0.94) was observed in HR+/HER2− group.

Figure 4.

Figure 4.

Open in a new tab

Diagram of image cropping for deep learning technique. The cubic shaped region-of-interest was selected at the largest cross-sectional area of the lesion and resized to 64 × 64 pixels. 18F-fluorodeoxyglucose (FDG) and apparent diffusion coefficient (ADC) images were obtained from positron emission tomography/computed tomography (PET/CT) and magnetic resonance imaging (MRI) scans, respectively. Baseline images were defined as PET0 and ADC0, respectively, and interim images were defined as PET1 and ADC1, respectively. Reprinted under a CC BY 4.0 license from: Choi, J.H., Kim, HA., Kim, W. et al. Early prediction of neoadjuvant chemotherapy response for advanced breast cancer using PET/MRI image deep learning. Sci Rep 10, 21149 (2020). https://doi.org/10.1038/s41598-020-77875-5

Based on the available evidence, the use of AI techniques could enable the early prediction of the response to PST even on pre-treatment PET/MRI examinations, whereas interim examinations seem to be the most effective for a “clinical assessment,” performed on the early variation of quantitative, mainly PET, parameters.

All that glitters is not gold: the dark shadows of AI

Preliminary evidence suggests a possible expanding role of AI to provide radiologists and clinicians with information that is not readily available from conventional images, such as molecular expression, response to treatment, and prognosis. However, there are still no robust data to suggest its imminent use in clinical practice. Indeed, as several studies highlighted, there is a huge heterogeneity in the applied AI methods, to the extent that guidelines and recommendations have been released in a bid to standardize image acquisition, processing, and analysis [27,49,50]. The need to assess the generalizability and robustness of the developed algorithms through external validation has also been pointed out. Once these methodological issues will be solved, the discussion of medico-legal implications of the clinical use of AI software is necessary for clinical adaption of radiomic tools. Indeed, clinicians may be responsible for decisions they did not make or the AI system may have a higher diagnostic accuracy than radiologists [51].

Concluding remarks

AI-empowered PET and MRI will play a central role in the assessment of patients with breast cancer. The ideal clinical scenario would be the “one-stop-shop” protocol to become a reality, while providing a set of comprehensive information from diagnosis to clinical management. It should be considered that other breast imaging modalities, such as ultrasound and mammography, will remain the backbone of breast imaging, especially considering their lower costs compared to MRI and PET. However, in the oncologic setting, whole-body imaging modalities like PET are preferred for their ability to detect distant metastasis, and MRI is the imaging modality with the highest sensitivity for the detection of breast cancer. Furthermore, both techniques represent an objective tool for tumor staging, treatment monitoring, and response assessment. Regarding functional and biological information that can be obtained through AI application, each modality seems most promising in a specific task. Indeed, while MRI has extensively been applied for breast cancer molecular subtyping and assessment of axillary lymph node involvement, PET shows high sensitivity in the early prediction of the response to PST, both using pre-treatment PET-derived features or the early variation of quantitative metabolic parameters at interim examinations. Additionally, several investigations demonstrated a significant improvement in the prediction task when clinical data are also added to the model. This combination of clinical and imaging data supports the idea that a comprehensive information package, including PET, MRI, and clinical information may further empower the performance of AI systems. However, a great deal remains to be done, in terms of standardization procedures and models validation. A major obstacle is also represented by the limited availability of hybrid PET/MRI scanners and their high related costs. Actions are being taken to overcome these issues, while providing recommendations and building public datasets. In conclusion, AI studies on PET/MRI systems in the breast cancer field are strongly encouraged, particularly in a multicenter setting, in an attempt to increase the robustness of the developed models and the standardization of MRI and PET imaging biomarkers.

Key points.

  1. PET/MRI can be empowered by radiomics and AI to meet the new needs of clinical and surgical oncology in breast cancer management.

  2. Preoperative and non-invasive assessment of tumor molecular profile, lymph node spread, recurrence risk, and early response to PST are the main goals of AI-enhanced diagnostic imaging.

  3. Standardization of AI methods and models’ validation are the essential prerequisites for their clinical implementation.

Reference

  • [1].Bray F, Ferlay J, Soerjomataram I, Siegel RL, Torre LA, Jemal A, Global cancer statistics 2018: GLOBOCAN estimates of incidence and mortality worldwide for 36 cancers in 185 countries., CA. Cancer J. Clin (2018). 10.3322/caac.21492. [DOI] [PubMed] [Google Scholar]
  • [2].Kalli S, Semine A, Cohen S, Naber SP, Makim SS, Bahl M, American Joint Committee on Cancer’s Staging System for Breast Cancer, Eighth Edition: What the Radiologist Needs to Know, RadioGraphics. 38 (2018) 1921–1933. 10.1148/rg.2018180056. [DOI] [PubMed] [Google Scholar]
  • [3].Padhani AR, Miles KA, Multiparametric imaging of tumor response to therapy, Radiology. (2010). 10.1148/radiol.10091760. [DOI] [PubMed] [Google Scholar]
  • [4].García-Figueiras R, Baleato-González S, Padhani AR, Luna-Alcalá A, Vallejo-Casas JA, Sala E, Vilanova JC, Koh D-M, Herranz-Carnero M, Vargas HA, How clinical imaging can assess cancer biology, Insights Imaging. 10 (2019) 28. 10.1186/s13244-019-0703-0. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • [5].Bruckmann NM, Morawitz J, Fendler WP, Ruckhäberle E, Bittner A-K, Giesel FL, Herrmann K, Antoch G, Umutlu L, Kirchner J, A Role of PET/MR in Breast Cancer?, Semin. Nucl. Med 52 (2022) 611–618. 10.1053/j.semnuclmed.2022.01.003. [DOI] [PubMed] [Google Scholar]
  • [6].Romeo V, Helbich TH, Pinker K, Breast PETMRI Hybrid Imaging and Targeted Tracers, J. Magn. Reson. Imaging (2022). 10.1002/jmri.28431. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • [7].Fowler AM, Strigel RM, Clinical advances in PET–MRI for breast cancer, Lancet Oncol. 23 (2022) e32–e43. 10.1016/S1470-2045(21)00577-5. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • [8].Gillies RJ, Kinahan PE, Hricak H, Radiomics: Images Are More than Pictures, They Are Data, Radiology. 278 (2016) 563–577. 10.1148/radiol.2015151169. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • [9].Shah SM, Khan RA, Arif S, Sajid U, Artificial intelligence for breast cancer analysis: Trends and directions, Comput. Biol. Med 142 (2022) 105221. 10.1016/j.compbiomed.2022.105221. [DOI] [PubMed] [Google Scholar]
  • [10].Reig B, Heacock L, Geras KJ, Moy L, Machine learning in breast MRI, J. Magn. Reson. Imaging 52 (2020) 998–1018. 10.1002/jmri.26852. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • [11].Lo Gullo R, Daimiel I, Morris EA, Pinker K, Combining molecular and imaging metrics in cancer: radiogenomics, Insights Imaging. 11 (2020) 1. 10.1186/s13244-019-0795-6. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • [12].Kohli M, Prevedello LM, Filice RW, Geis JR, Implementing machine learning in radiology practice and research, Am. J. Roentgenol (2017). 10.2214/AJR.16.17224. [DOI] [PubMed] [Google Scholar]
  • [13].Polyak K, Heterogeneity in breast cancer, J. Clin. Invest 121 (2011) 3786–3788. 10.1172/JCI60534. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • [14].Cleator S, Ashworth A, Molecular profiling of breast cancer: clinical implications, Br. J. Cancer 90 (2004) 1120–1124. 10.1038/sj.bjc.6601667. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • [15].Helbich TH, Matzek W, Fuchsjäger MH, Stereotactic and ultrasound-guided breast biopsy, Eur. Radiol 14 (2004) 383–393. 10.1007/s00330-003-2141-z. [DOI] [PubMed] [Google Scholar]
  • [16].Marino MA, Avendano D, Zapata P, Riedl CC, Pinker K, Lymph Node Imaging in Patients with Primary Breast Cancer: Concurrent Diagnostic Tools, Oncologist. (2020). 10.1634/theoncologist.2019-0427. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • [17].Thompson JL, Wright GP, Contemporary approaches to the axilla in breast cancer, Am. J. Surg (2022). 10.1016/j.amjsurg.2022.11.036. [DOI] [PubMed] [Google Scholar]
  • [18].Noguchi M, Inokuchi M, Yokoi-Noguchi M, Morioka E, Haba Y, Conservative axillary surgery is emerging in the surgical management of breast cancer, Breast Cancer. 30 (2023) 14–22. 10.1007/s12282-022-01409-2. [DOI] [PubMed] [Google Scholar]
  • [19].Angarita FA, Brumer R, Castelo M, Esnaola NF, Edge SB, Takabe K, De-Escalating the Management of In Situ and Invasive Breast Cancer, Cancers (Basel). 14 (2022) 4545. 10.3390/cancers14194545. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • [20].Lyman GH, Temin S, Edge SB, Newman LA, Turner RR, Weaver DL, Benson AB, Bosserman LD, Burstein HJ, Cody H, Hayman J, Perkins CL, Podoloff DA, Giuliano AE, Sentinel Lymph Node Biopsy for Patients With Early-Stage Breast Cancer: American Society of Clinical Oncology Clinical Practice Guideline Update, J. Clin. Oncol 32 (2014) 1365–1383. 10.1200/JCO.2013.54.1177. [DOI] [PubMed] [Google Scholar]
  • [21].Romeo V, Accardo G, Perillo T, Basso L, Garbino N, Nicolai E, Maurea S, Salvatore M, Assessment and Prediction of Response to Neoadjuvant Chemotherapy in Breast Cancer: A Comparison of Imaging Modalities and Future Perspectives, Cancers (Basel). 13 (2021) 3521. 10.3390/cancers13143521. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • [22].Varsanik MA, Shubeck SP, De-Escalating Breast Cancer Therapy, Surg. Clin. North Am 103 (2023) 83–92. 10.1016/j.suc.2022.08.005. [DOI] [PubMed] [Google Scholar]
  • [23].Spring LM, Bar Y, Isakoff SJ, The Evolving Role of Neoadjuvant Therapy for Operable Breast Cancer, J. Natl. Compr. Cancer Netw 20 (2022) 723–734. 10.6004/jnccn.2022.7016. [DOI] [PubMed] [Google Scholar]
  • [24].De La Cruz LM, Harhay MO, Zhang P, Ugras S, Impact of Neoadjuvant Chemotherapy on Breast Cancer Subtype: Does Subtype Change and, if so, How?, Ann. Surg. Oncol 25 (2018) 3535–3540. 10.1245/s10434-018-6608-1. [DOI] [PubMed] [Google Scholar]
  • [25].Kittaneh M, Montero AJ, Glück S, Molecular Profiling for Breast Cancer: A Comprehensive Review, Biomark. Cancer 5 (2013) BIC.S9455. 10.4137/BIC.S9455. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • [26].Goncalves R, Bose R, Using Multigene Tests to Select Treatment for Early-Stage Breast Cancer, J. Natl. Compr. Cancer Netw 11 (2013) 174–182. 10.6004/jnccn.2013.0025. [DOI] [PubMed] [Google Scholar]
  • [27].Stanzione A, Cuocolo R, Ugga L, Verde F, Romeo V, Brunetti A, Maurea S, Oncologic Imaging and Radiomics: A Walkthrough Review of Methodological Challenges, Cancers (Basel). 14 (2022) 4871. 10.3390/cancers14194871. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • [28].Afshar P, Mohammadi A, Plataniotis KN, Oikonomou A, Benali H, From Handcrafted to Deep-Learning-Based Cancer Radiomics: Challenges and Opportunities, IEEE Signal Process. Mag 36 (2019) 132–160. 10.1109/MSP.2019.2900993. [DOI] [Google Scholar]
  • [29].Reig B, Heacock L, Geras KJ, Moy L, Machine learning in breast MRI, J. Magn. Reson. Imaging (2020). 10.1002/jmri.26852. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • [30].Huang S, Franc BL, Harnish RJ, Liu G, Mitra D, Copeland TP, Arasu VA, Kornak J, Jones EF, Behr SC, Hylton NM, Price ER, Esserman L, Seo Y, Exploration of PET and MRI radiomic features for decoding breast cancer phenotypes and prognosis, Npj Breast Cancer. 4 (2018) 24. 10.1038/s41523-018-0078-2. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • [31].Romeo V, Clauser P, Rasul S, Kapetas P, Gibbs P, Baltzer PAT, Hacker M, Woitek R, Helbich TH, Pinker K, AI-enhanced simultaneous multiparametric 18F-FDG PET/MRI for accurate breast cancer diagnosis, Eur. J. Nucl. Med. Mol. Imaging 49 (2022) 596–608. 10.1007/s00259-021-05492-z. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • [32].Romeo V, Kapetas P, Clauser P, Baltzer PAT, Rasul S, Gibbs P, Hacker M, Woitek R, Pinker K, Helbich TH, A Simultaneous Multiparametric 18F-FDG PET/MRI Radiomics Model for the Diagnosis of Triple Negative Breast Cancer, Cancers (Basel). 14 (2022) 3944. 10.3390/cancers14163944. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • [33].Umutlu L, Kirchner J, Bruckmann NM, Morawitz J, Antoch G, Ingenwerth M, Bittner A-K, Hoffmann O, Haubold J, Grueneisen J, Quick HH, Rischpler C, Herrmann K, Gibbs P, Pinker-Domenig K, Multiparametric Integrated 18F-FDG PET/MRI-Based Radiomics for Breast Cancer Phenotyping and Tumor Decoding, Cancers (Basel). 13 (2021) 2928. 10.3390/cancers13122928. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • [34].Zhang J, Li L, Zhe X, Tang M, Zhang X, Lei X, Zhang L, The Diagnostic Performance of Machine Learning-Based Radiomics of DCE-MRI in Predicting Axillary Lymph Node Metastasis in Breast Cancer: A Meta-Analysis, Front. Oncol 12 (2022) 1–10. 10.3389/fonc.2022.799209. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • [35].Chen K, Yin G, Xu W, Predictive Value of 18F-FDG PET/CT-Based Radiomics Model for Occult Axillary Lymph Node Metastasis in Clinically Node-Negative Breast Cancer, Diagnostics. 12 (2022) 997. 10.3390/diagnostics12040997. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • [36].Cheng J, Ren C, Liu G, Shui R, Zhang Y, Li J, Shao Z, Development of High-Resolution Dedicated PET-Based Radiomics Machine Learning Model to Predict Axillary Lymph Node Status in Early-Stage Breast Cancer, Cancers (Basel). 14 (2022) 950. 10.3390/cancers14040950. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • [37].Morawitz J, Sigl B, Rubbert C, Bruckmann N-M, Dietzel F, Häberle LJ, Ting S, Mohrmann S, Ruckhäberle E, Bittner A-K, Hoffmann O, Baltzer P, Kapetas P, Helbich T, Clauser P, Fendler WP, Rischpler C, Herrmann K, Schaarschmidt BM, Stang A, Umutlu L, Antoch G, Caspers J, Kirchner J, Clinical decision support for axillary lymph node staging in newly diagnosed breast cancer patients based on 18 F-FDG PET/MRI and machine-learning, J. Nucl. Med (2022) jnumed.122.264138. 10.2967/jnumed.122.264138. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • [38].Saha A, Harowicz MR, Wang W, Mazurowski MA, A study of association of Oncotype DX recurrence score with DCE-MRI characteristics using multivariate machine learning models, J. Cancer Res. Clin. Oncol (2018). 10.1007/s00432-018-2595-7. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • [39].Kim HJ, Choi WJ, Kim HH, Cha JH, Shin HJ, Chae EY, Association between Oncotype DX recurrence score and dynamic contrast-enhanced MRI features in patients with estrogen receptor-positive HER2-negative invasive breast cancer, Clin. Imaging 75 (2021) 131–137. 10.1016/j.clinimag.2021.01.021. [DOI] [PubMed] [Google Scholar]
  • [40].Jacobs MA, Umbricht CB, Parekh VS, El Khouli RH, Cope L, Macura KJ, Harvey S, Wolff AC, Integrated Multiparametric Radiomics and Informatics System for Characterizing Breast Tumor Characteristics with the OncotypeDX Gene Assay, Cancers (Basel). 12 (2020) 2772. 10.3390/cancers12102772. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • [41].Romeo V, Cuocolo R, Sanduzzi L, Carpentiero V, Caruso M, Lama B, Garifalos D, Stanzione A, Maurea S, Brunetti A, MRI Radiomics and Machine Learning for the Prediction of Oncotype Dx Recurrence Score in Invasive Breast Cancer, Cancers (Basel). 15 (2023) 1840. 10.3390/cancers15061840. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • [42].Ha R, Chang P, Mutasa S, Karcich J, Goodman S, Blum E, Kalinsky K, Liu MZ, Jambawalikar S, Convolutional Neural Network Using a Breast MRI Tumor Dataset Can Predict Oncotype Dx Recurrence Score, J. Magn. Reson. Imaging 49 (2019) 518–524. 10.1002/jmri.26244. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • [43].O’Donnell JPM, Gasior SA, Davey MG, O’Malley E, Lowery AJ, McGarry J, O’Connell AM, Kerin MJ, McCarthy P, The accuracy of breast MRI radiomic methodologies in predicting pathological complete response to neoadjuvant chemotherapy: A systematic review and network meta-analysis, Eur. J. Radiol 157 (2022) 110561. 10.1016/j.ejrad.2022.110561. [DOI] [PubMed] [Google Scholar]
  • [44].Pesapane F, Agazzi GM, Rotili A, Ferrari F, Cardillo A, Penco S, Dominelli V, D’Ecclesiis O, Vignati S, Raimondi S, Bozzini A, Pizzamiglio M, Petralia G, Nicosia L, Cassano E, Prediction of the Pathological Response to Neoadjuvant Chemotherapy in Breast Cancer Patients With MRI-Radiomics: A Systematic Review and Meta-analysis, Curr. Probl. Cancer 46 (2022) 100883. 10.1016/j.currproblcancer.2022.100883. [DOI] [PubMed] [Google Scholar]
  • [45].Sollini M, Cozzi L, Ninatti G, Antunovic L, Cavinato L, Chiti A, Kirienko M, PET/CT radiomics in breast cancer: Mind the step, Methods. (2020). 10.1016/j.ymeth.2020.01.007. [DOI] [PubMed] [Google Scholar]
  • [46].Li P, Wang X, Xu C, Liu C, Zheng C, Fulham MJ, Feng D, Wang L, Song S, Huang G, 18F-FDG PET/CT radiomic predictors of pathologic complete response (pCR) to neoadjuvant chemotherapy in breast cancer patients, Eur. J. Nucl. Med. Mol. Imaging (2020). 10.1007/s00259-020-04684-3. [DOI] [PubMed] [Google Scholar]
  • [47].Choi JH, Kim H-A, Kim W, Lim I, Lee I, Byun BH, Noh WC, Seong M-K, Lee S-S, Il Kim B, Choi CW, Lim SM, Woo S-K, Early prediction of neoadjuvant chemotherapy response for advanced breast cancer using PET/MRI image deep learning, Sci. Rep 10 (2020) 21149. 10.1038/s41598-020-77875-5. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • [48].Umutlu L, Kirchner J, Bruckmann N-M, Morawitz J, Antoch G, Ting S, Bittner A-K, Hoffmann O, Häberle L, Ruckhäberle E, Catalano OA, Chodyla M, Grueneisen J, Quick HH, Fendler WP, Rischpler C, Herrmann K, Gibbs P, Pinker K, Multiparametric 18F-FDG PET/MRI-Based Radiomics for Prediction of Pathological Complete Response to Neoadjuvant Chemotherapy in Breast Cancer, Cancers (Basel). 14 (2022) 1727. 10.3390/cancers14071727. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • [49].Mongan J, Moy L, Kahn CE, Checklist for Artificial Intelligence in Medical Imaging (CLAIM): A Guide for Authors and Reviewers, Radiol. Artif. Intell 2 (2020) e200029. 10.1148/ryai.2020200029. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • [50].Lambin P, Leijenaar RTH, Deist TM, Peerlings J, de Jong EEC, van Timmeren J, Sanduleanu S, Larue RTHM, Even AJG, Jochems A, van Wijk Y, Woodruff H, van Soest J, Lustberg T, Roelofs E, van Elmpt W, Dekker A, Mottaghy FM, Wildberger JE, Walsh S, Radiomics: the bridge between medical imaging and personalized medicine, Nat. Rev. Clin. Oncol 14 (2017) 749–762. 10.1038/nrclinonc.2017.141. [DOI] [PubMed] [Google Scholar]
  • [51].Carter SM, Rogers W, Than K, Frazer H, Richards B, Houssami N, The ethical, legal and social implications of using artificial intelligence systems in breast cancer care, The Breast. 49 (2020) 25–32. 10.1016/j.breast.2019.10.001. [DOI] [PMC free article] [PubMed] [Google Scholar]

RESOURCES