Radiomics and artificial intelligence for predicting response to neoadjuvant drug therapy in patients with breast cancer: a review

Full Text

INTRODUCTION

Breast cancer is the most common malignant neoplasm among women worldwide. In 2020, 7.8 million women were living with a breast cancer diagnosis made within the previous 5 years, with 2.3 million new cases and 685,000 deaths reported globally [1, 2].

The primary goal in managing patients with a breast cancer diagnosis is to select the most effective and economically feasible therapeutic protocol, taking into account the tumor’s molecular subtype and the patient’s individual response to treatment [3].

There are two main treatment approaches for breast cancer:

• Local therapy, including surgery and radiation therapy;
• Systemic therapy, including chemotherapy, endocrine (hormone) therapy, and targeted therapy.

Selection of the optimal treatment strategy for each patient depends on many factors, including age, menopausal status, molecular subtype, tumor stage, overall health, and patient preferences. Systemic therapy may be given after surgery in the adjuvant setting, whereas neoadjuvant therapy precedes surgery [4, 5].

Neoadjuvant chemotherapy (NACT) is the standard of care for locally advanced, initially inoperable invasive breast cancer, an aggressive form characterized by tumor size greater than 5 cm and potential involvement of the skin or chest wall [3]. In addition, according to the guidelines of the Russian Society of Oncomammologists, the Association of Oncologists of Russia, the clinical guidelines for breast cancer treatment of the Ministry of Health of the Russian Federation, as well as the American Society of Clinical Oncology and the National Comprehensive Cancer Network, NACT is considered the preferred first-line treatment. It is indicated for patients with initially operable triple-negative and HER2-positive (human epidermal growth factor receptor 2–positive) breast cancer, facilitates favorable treatment responses in approximately 30% of women with aggressive disease subtypes, and reduces recurrence rates by 50% [6].

The main objectives of NACT are:

• to reduce tumor size to optimize the surgical treatment;
• to assess the effectiveness and in vivo chemosensitivity of the tumor, enabling timely modification of the treatment strategy;
• to obtain prognostically relevant information based on the degree of the pathologic treatment response (partial or complete pathologic response, pPR or pCR, respectively), thereby facilitating optimization and selection of adjuvant therapy tactics [3].

Nevertheless, despite the recognized advantages of NACT, heterogeneous treatment responses among patients remain a serious clinical challenge. The same chemotherapeutic agents are used in NACT regimens as in adjuvant therapy, which results in the development of comparable short- and long-term adverse effects. These include fatigue, vomiting, nausea, cognitive impairment, alopecia, infertility, osteoporosis, cardiomyopathy, immunosuppression, infectious complications, leukemia, neuropathy, and others. In addition, the magnitude of the pathologic response depends on the molecular subtype of breast cancer. Notably, up to 30% of patients with breast cancer derive no clinical benefit from NACT while still being exposed to its toxic and other adverse effects [7].

The degree of pathologic response is currently the only reliable biomarker of disease-free and overall survival. pCR has been shown to be strongly associated with long-term survival and an approximately 80% reduction in recurrence risk, as demonstrated in the I-SPY 2 trial (Investigation of Serial Studies to Predict Your Therapeutic Response with Imaging and Molecular Analysis 2) [8].

In breast cancer, pCR is defined as the absence of residual invasive tumor cells in both the primary tumor site and the regional lymph nodes. Previous studies have shown that patients with triple-negative and HER2-positive breast cancer are more likely to achieve pCR [9, 10].

However, the degree of pathologic response can be reliably assessed only at the time of surgical treatment [11]. This creates a clear need for the development of noninvasive predictive models capable of identifying patients who are likely to benefit from NACT and those for whom this treatment will be ineffective. Such models may be developed using radiomics and artificial intelligence techniques, which in turn could optimize healthcare costs and improve the therapeutic efficacy in patients with breast cancer.

This review analyzes published data on the potential to improve the prognostic value of imaging techniques for evaluating the pathologic response to NACT in patients with breast cancer using radiomics and artificial intelligence.

SEARCH METHODOLOGY

We conducted a scientific data search using the bibliographic and analytical databases PubMed, Google Scholar, and Scopus. A total of 58 publications from 2006 to 2024 were included in the analysis. The search was performed using the following keywords: breast cancer / рак молочной железы, neoadjuvant chemotherapy / неоадъювантная химиотерапия, mammography / маммография, ultrasound / ультразвуковое исследование, magnetic resonance imaging / магнитно-резонансная томография, complete pathological response / полный патоморфологический ответ, radiomics / радиомика.

RADIOMICS AND ARTIFICIAL INTELLIGENCE

Radiomics is a rapidly evolving field of medicine focused on exploring the relationships between qualitative and quantitative information derived from medical images. The analysis of quantitative features enables the construction of multidimensional models, which, when combined with clinical data, support medical decision-making [12]. Given the large volume of data generated by radiomic analysis, artificial intelligence has become an important tool for their in-depth and comprehensive evaluation, which is especially relevant in the context of personalized medicine aimed at individualized patient treatment [13–14].

Radiomic analysis of medical images includes several stages:

• image acquisition
• image segmentation
• feature extraction and selection
• model development.

Pixel intensity values in imaging modalities such as magnetic resonance imaging (MRI), ultrasound (US) imaging, and mammography are subject to substantial variability depending on imaging parameters and do not always correlate with the physical properties of tissues, in contrast to computed tomography, where the Hounsfield scale is used for quantitative assessment of tissue density. To improve the stability and reproducibility of results, the use of the same scanners and acquisition protocols is recommended; when this is not feasible, data harmonization techniques should be applied [15–17].

Image segmentation, which enables the delineation of regions of interest, represents the most critical stage of radiomic analysis. This process can be performed manually, semi-automatically, or fully automatically. Manual segmentation may introduce subjective bias, as many radiomic features are sensitive to intra- and interobserver variability during region-of-interest delineation. Therefore, studies employing manual segmentation must thoroughly assess the reproducibility of the extracted features and exclude non-reproducible parameters from further analysis [18, 19].

Semi-automatic segmentation generally demonstrates good performance for homogeneous tumors, whereas heterogeneous lesions often require substantial manual correction [17]. Fully automated segmentation based on deep learning is actively developing; in particular, there are models capable of segmenting images of various organs. The main limitation of automatic segmentation is its frequent lack of reproducibility when applied to another dataset [12, 19, 20].

Radiomic feature extraction involves the calculation of various mathematical descriptors to quantitatively characterize grayscale intensity levels within each region of interest. Numerous methods and formulas exist for their computation, and to enhance data reproducibility, adherence to the standards of the Image Biomarker Standardization Initiative (IBSI) is recommended [12, 19]. These features can be classified into four main groups:

• Shape features that can be used to describe the geometric properties of the delineated region of interest, including maximum linear dimension, volume, surface area, and boundary characteristics;
• First-order features that can be used to describe the distribution of individual voxel values without considering their spatial relationships and include the mean, standard deviation, variance, skewness, kurtosis, and entropy;
• Second-order features that are based on the calculation of statistical relationships between neighboring voxels, reflecting the spatial distribution of intensity and structural heterogeneity.

Higher-order radiomic features are generated using filters and mathematical transformations for subsequent analysis [12, 19].

At the next stage, feature selection is performed to exclude irrelevant and non-reproducible data. This can be achieved using both statistical methods and machine- or deep-learning approaches. Initially, all extracted features are included, followed by a preliminary analysis to identify the most stable and reproducible ones, facilitating subsequent reduction in their number through correlation and redundancy analysis [17].

The remaining non-correlated yet informative features can then be used as input variables for model construction aimed at solving specific tasks, such as those designed to differentiate malignant from benign lesions. These models are typically developed by splitting the data into training and testing sets, whereas the most robust models are additionally validated on external datasets to ensure reproducibility of the obtained results [18, 21].

PROSPECTS FOR USING RADIOMICS AND ARTIFICIAL INTELLIGENCE IN PREDICTING RESPONSE TO NEOADJUVANT CHEMOTHERAPY IN PATIENTS WITH BREAST CANCER

Mammography and Contrast-Enhanced Spectral Mammography

Mammography is the gold standard for imaging pathologic changes in the breast. This method is characterized by a low radiation dose and is used both for screening and for the diagnosis of mass lesions, architectural distortion, and microcalcifications, with an accuracy ranging from 85% to 90% [22, 23]. Mammography was one of the first imaging modalities to incorporate artificial intelligence technologies, beginning with traditional computer-aided diagnosis (CAD) systems. CAD systems for mammography have been in clinical use for more than 10 years, providing a reference standard against which newer machine-learning and deep-learning methods can be compared [24].

Several artificial intelligence studies have demonstrated the potential to assess early response to NACT using baseline mammographic images (Supplement 1) [25, 26]. Shin et al. [25], for example, applied an Image Pyramid With Multiple Scales method to mammographic data analysis. This approach is used to extract texture and other features from medical images (in particular, mammograms), enabling their inclusion in radiomic workflows. The authors developed a radiomics-based model (incorporating texture features and tumor shape descriptors) that showed good prognostic performance for predicting pCR, with an area under the curve (AUC), accuracy, sensitivity, and specificity of 0.803, 0.75, 0.733, and 0.767, respectively. Use of a multiscale image pyramid as a preprocessing technique improves feature extraction by allowing images to be analyzed at different levels of detail. In this way, the method can serve as a radiomic analysis tool, facilitating the extraction of more informative and relevant features for subsequent evaluation and predictive model development.

Skarping et al. [26] were the first to apply artificial intelligence techniques for the analysis of mammographic images and prediction of treatment response. Using deep learning methods, they developed a model for the automated analysis of digital mammograms to assess response to NACT in breast cancer. The model demonstrated good predictive performance for pCR, with an AUC of 0.71 (95% confidence interval [CI], 0.53–0.90), whereas sensitivity and specificity were 46% and 90%, respectively. It should be noted that deep learning methods are being increasingly incorporated into radiomics for the automated extraction and analysis of relationships between quantitative features of medical images.

Contrast-enhanced spectral mammography (CESM) is a novel breast imaging technique based on dual-energy acquisition, in which low-energy and high-energy images obtained after the administration of an iodine-based contrast agent are used to generate diagnostic images [27, 28]. CESM may be used as an alternative to breast MRI in patients with contraindications, such as severe claustrophobia or the presence of incompatible implants (pacemakers, defibrillators, neurostimulators, cochlear implants) [29, 30]. In multiple studies, CESM has demonstrated diagnostic accuracy and sensitivity comparable to those of MRI. This method is used for breast cancer screening, diagnosis, tumor staging, monitoring, and prediction of response to NACT [27, 28, 31].

Xing et al. [32] performed a quantitative analysis of gray-level values derived from CESM for the early prediction of pathologic response to NACT in breast cancer. The authors applied the statistical t test to evaluate the percentage reduction in the gray value of CESM subtraction images (ΔCGV) in the craniocaudal and mediolateral projections. The diagnostic model was based on the determination of a threshold value that separated predictive outcomes for patients with and without treatment response. The results indicated that the gray values of CESM subtraction images in patients with pCR were significantly lower than in those with non-pCR. In addition, a significant difference in ΔCGV between the two response categories was observed after the second cycle of NACT. Specifically, ΔCGV was higher in the pCR group than in the non-pCR group (p < 0.001). Thus, this parameter in the craniocaudal and mediolateral projections demonstrated prognostic value for predicting response to NACT, as confirmed by AUC values of 0.776 and 0.733, respectively. At a threshold value of >26.41 in the craniocaudal projection, sensitivity and specificity reached 75% and 72.15%, respectively. For ΔCGV in the mediolateral projection, at a threshold value of >13.59, sensitivity and specificity were 81.25% and 51.90%, respectively. However, it should be noted that the use of a single parameter alone, namely, the gray-level value, represents a considerable limitation of this study.

Wang et al. [33] applied radiomic analysis to evaluate CESM images to predict breast cancer insensitivity to NACT. To construct the nomogram, radiomic features as well as three independent clinical risk factors were used:

• background parenchymal enhancement (BPE)
• HER2 status
• Ki-67 index (a marker reflecting tumor cell proliferative activity).

Treatment response was assessed using the RECIST (Response Evaluation Criteria in Solid Tumors) by comparing the maximum tumor diameter before the initiation of NACT with the size of the residual lesion after its completion. Two patient groups were distinguished according to the course of the tumor process:

• Disease stabilization, defined as a tumor size reduction of <30% or an increase of <20%;
• Disease progression, defined as a tumor size increase of ≥20% (absence of response to NACT).

The proposed nomogram demonstrated high predictive performance for identifying breast cancers insensitive to NACT prior to treatment initiation, with an AUC of 0.810 (95% CI, 0.575–0.948), accuracy of 0.80, sensitivity of 0.90, and specificity of 0.70.

Mao et al. [34] used several machine learning algorithms to analyze quantitative features extracted from CESM images obtained both from the tumor itself and from adjacent regions, encompassing intratumoral and peritumoral areas. The authors noted that a model combining radiomic features extracted from both intratumoral and peritumoral regions (including margins up to 5 mm) demonstrated high predictive performance for assessing response to NACT (AUC, sensitivity, and specificity of 0.85, 0.58, and 0.91, respectively).

Ultrasound Imaging

US imaging is based on the detection of reflected echo signals from high-frequency sound waves. Quantitative US, which applies mathematical signal processing to obtain objective numerical tissue characteristics, has gained particular relevance, especially in the context of radiomic analysis. Moreover, the US has several advantages, including:

• Availability and relatively low cost;
• Independence from contrast agents;
• The possibility of frequent repeated examinations due to the absence of ionizing radiation.

Among the most commonly used radiomic US features for predicting response to NACT are:

• spectral slope (SS)
• spectral intercept at 0 MHz (SI)
• midband fit (MBF)
• average scatterer diameter (ASD)
• average acoustic concentration (AAC)
• attenuation coefficient estimate (ACE)
• spacing between acoustic scatterers (SAS).

In addition, machine learning methods such as linear discriminant analysis, k-nearest neighbors, and support vector machines are most commonly used to evaluate the effectiveness of NACT based on US data (Supplement 2) [35–38].

Sadeghi-Naini et al. [39] used linear discriminant analysis to evaluate the predictive value of midband fit, spectral slope, and spectral intercept at 0 MHz for predicting treatment response. They found that the best separation between patient groups (with and without response to NACT) was achieved through the combination of texture and spectral features derived from quantitative US parametric maps obtained after one week of treatment, as confirmed by the following performance indicators: 100% sensitivity and specificity and an AUC of 1.

Sannachi et al. [40] applied machine learning techniques to develop predictive models based on parametric and texture features extracted from quantitative US data to predict tumor response to NACT at 1, 4, and 8 weeks after treatment initiation. The support vector machine model demonstrated the highest predictive performance, with AUC values of 0.71, 0.87, and 0.92 at weeks 1, 4, and 8, respectively.

DiCenzo et al. [41] developed a k-nearest neighbors–based model that showed high predictive value for determining response to NACT, with an AUC of 0.73 and accuracy, sensitivity, and specificity of 87%, 91%, and 83%, respectively. The proposed prediction model incorporated three features: homogeneity of average acoustic concentration, spectral intercept energy, and energy of spacing between acoustic scatterers.

There is a study in which both radiomic features and molecular tumor characteristics, specifically estrogen receptor (ER+/−), progesterone receptor (PR+/−), and HER2+/− status, were used to construct clinico-diagnostic models [42]. Thus, Tadayyon et al. [42] conducted a prospective study in which a multiparametric model for predicting response to NACT was developed using machine learning methods (linear discriminant analysis, k-nearest neighbors, and support vector machines). The model was based on texture features extracted from US images combined with molecular tumor characteristics. The regions of interest on US images were delineated before treatment both in the central part of the tumor and along its margins (at depths of 3, 5, and 10 mm). The model constructed using the k-nearest neighbors algorithm, incorporating radiomic features (ROI in the tumor core plus margins up to 5 mm) and molecular markers, demonstrated the best performance, with an AUC of 0.81 and accuracy, sensitivity, and specificity of 88%, 90%, and 79%, respectively. However, the clinico-diagnostic model combining both radiomic and molecular markers showed a decrease in accuracy and AUC to 79% and 0.71, respectively.

There are also studies in which compression and shear-wave elastography are used to predict response to NACT. These imaging modalities assess the mechanical properties of tissues, such as stiffness and elasticity. Compression elastography involves evaluating tissue deformation after the application of static compression using a manual maneuver with an ultrasound transducer. In contrast, shear-wave elastography quantitatively measures the propagation velocity of shear waves within the tissue, which are induced by focused acoustic radiation force [43–45].

Thus, Fernandes et al. [45] performed compression elastography to calculate relative changes in the strain ratio within breast tumor tissue during the treatment course. The strain ratio demonstrated a significant difference between the two response groups after 2 weeks of NACT (p < 0.01). The authors applied machine learning methods to evaluate the prognostic performance of the relative change in the strain ratio as a marker of response to NACT. A model based on a naive Bayes classifier predicted pCR with a sensitivity of 84% and a specificity of 85%.

Ma et al. [43] developed a multivariate linear regression model and demonstrated that a combined assessment of the Ki-67 index and relative changes in shear-wave elastography parameters after the second cycle of NACT had high prognostic value. The AUC for the Ki-67 index and for the relative change in stiffness in identifying nonresponders was 0.84 and 0.82, respectively, whereas their combination achieved an AUC of 0.93.

Gu et al. [46] likewise reported that a combined evaluation of the Ki-67 index and shear-wave elastography parameters improved the predictive performance of the model at mid-treatment, reaching an AUC of 0.80. In addition, a novel radiomic biomarker, the mass characteristic frequency, was identified, the change of which is associated with a specific behavior or response of tumor tissue to applied mechanical stimulation.

Magnetic Resonance Imaging with Dynamic Contrast Enhancement

Dynamic contrast–enhanced MRI requires intravenous administration of a contrast agent to obtain functional (over time) tissue information. This imaging modality provides essential data on tumor morphology, including size, shape characteristics, and texture heterogeneity, as well as features of tumor vascularization. Dynamic contrast–enhanced MRI can be used in the following settings (Supplement 3):

• Screening of women at high risk for breast cancer;
• Tumor detection and staging;
• Assessment of treatment efficacy and early prediction of response to NACT [47].

Pesapane et al. [48] conducted a systematic review and meta-analysis of 43 and 34 studies, respectively, published between 2013 and 2021, in which radiomic analysis of breast MRI was used to predict pCR in patients with breast cancer undergoing NACT. In addition, the authors evaluated the methodological quality of these studies using the Radiomics Quality Score (RQS). RQS is a parameter used to assess the quality of radiomics research. This parameter enables standardized and systematic evaluation of methodological aspects of radiomics research, such as study design, validation, data handling, and result analysis. The systematic review was conducted in accordance with the PRISMA-DTA (Preferred Reporting Items for Systematic Reviews and Meta-Analysis for Diagnostic Test Accuracy) guidelines. The pooled AUC was 0.78 (95% CI, 0.74–0.81). Heterogeneity according to the I² statistic was substantial (71.05%, p < 0.001), indicating significant variability between studies. The mean RQS was 12.9 points (range: −1 to 26), corresponding to 36% of the maximum achievable score. It was established that year of publication, magnetic field strength, and total RQS did not explain the observed heterogeneity, indicating the need for a more in-depth analysis of the sources of these differences. Furthermore, all included studies were single-center investigations, which affects reproducibility and standardization of radiomic models and limits their broad implementation for routine clinical assessment of response to NACT.

In some studies, various combinations of molecular and radiomic features were also tested to predict treatment response either before or at an early stage of NACT [49–53].

Jimenez et al. [49] developed a prognostic model based on radiomic features extracted from pretreatment MRI images and tumor-infiltrating lymphocyte (TIL) levels in biopsy specimens. The authors hypothesized that patients with TIL levels >20% and a radiomic signature value <0.33 would achieve pCR. The combined prognostic model demonstrated superior diagnostic performance, with an AUC of 0.752 and accuracy, sensitivity, and specificity of 83%, 56%, and 97%, respectively.

Jahani et al. [50] analyzed dynamic contrast–enhanced MRI images to assess changes in intratumoral heterogeneity. The authors identified two distinct categories of features. The first category included:

• Voxel-wise tumor deformation characteristics, reflecting changes in tumor size, orientation, and shape (including the Jacobian parameter, defined as the ratio of tumor volume after the first cycle of NACT to baseline tumor volume);
• The anisotropic deformation index (ADI);
• The slab–rod index (SRI).

The second category comprised voxel-based changes in dynamic contrast–enhanced features, including:

• Peak enhancement (PE);
• Wash-in slope (WIS);
• Washout slope (WOS);
• Signal enhancement ratio (SER).

The voxel-based model constructed using logistic regression demonstrated the best performance for pCR prediction, with an AUC of 0.74. In addition, the potential of demographic and molecular features for predicting pCR and recurrence-free survival was evaluated. The pCR prediction model incorporating age, race, hormone receptor status, and functional tumor volume achieved an AUC of 0.71, whereas the inclusion of demographic, molecular, and voxel-based features increased the AUC to 0.78. The authors also compared the prognostic performance of voxel-based and dynamic features; however, neither model demonstrated superior performance (AUC = 0.71, p > 0.05). Thus, voxel-based features exhibited greater diagnostic value for pCR prediction.

Sutton et al. [51] applied the random forest algorithm to develop a pCR prediction model based on radiomic features extracted from pre- and post-treatment MRI images and molecular tumor characteristics. Three models were constructed:

• Model 1, based solely on radiomic features (AUC, sensitivity, and specificity: 0.83, 0.77, and 0.69, respectively);
• Model 2, based on molecular tumor subtype and radiomic features, demonstrating a slight improvement in prognostic performance (AUC, sensitivity, and specificity: 0.78, 0.79, and 0.69, respectively);
• Model 3, based on radiomic features excluding contrast enhancement intensity parameters on MRI, showing performance comparable to that of the first model (AUC, sensitivity, and specificity: 0.78, 0.79, and 0.69, respectively).

Fan et al. [52] evaluated changes in tumor heterogeneity using texture analysis of MRI images acquired before and after two cycles of NACT. A support vector machine was used to construct the prediction model. The prognostic models based on pretreatment radiomic features and Jacobian map parameters yielded AUC values of 0.568 and 0.630, respectively. In contrast, the use of radiomic features extracted from MRI images obtained after the second cycle of NACT improved the predictive performance of the model, as evidenced by an AUC of 0.77. In addition, the model based on the assessment of feature changes between pre-treatment and on-treatment images demonstrated an AUC of 0.73. The combined model incorporating radiomic features and molecular tumor subtype data showed the highest prognostic value, with AUC, sensitivity, and specificity of 0.81, 0.83, and 0.80, respectively.

Hussain et al. [53] developed pCR prediction models using machine learning methods by combining radiomic features extracted from MRI images with molecular tumor subtype data and the Ki-67 index. The model based on molecular tumor subtype, constructed using a boosted tree ensemble with random undersampling, demonstrated an AUC of 0.82 and an accuracy of 0.84. Texture features extracted from MRI images obtained before treatment, after the first cycle, and at mid-treatment showed the following diagnostic performance: AUC values of 0.88, 0.72, and 0.78, and accuracy values of 0.86, 0.82, and 0.76, respectively. Combining features from two time points (baseline and after the first cycle of NACT) resulted in high predictive performance, with an AUC of 0.96 and an accuracy of 0.84. Moreover, the addition of molecular tumor subtype data further enhanced the prognostic performance, increasing the AUC and accuracy to 0.98 and 0.94, respectively [53].

Parametric Response Maps have been used to investigate regions of increased and decreased intratumoral signal intensity at an early stage of treatment [54, 55]. Thus, Cho et al. [54] applied the t test (Student’s t test) to compare the predictive performance of conventional pharmacokinetic parameters, namely the volume transfer constant (K_trans), the rate constant (k_ep), and the extravascular extracellular volume fraction (v_e), with analysis based on parametric response maps. Their assessment included voxel-wise comparison between dynamic contrast-enhanced MRI images obtained before treatment and after the first cycle of NACT. Voxels showing increased (>10%) and decreased signal intensity were labeled as PRM_SI+ and PRM_SI−, respectively. As a result, the authors found no significant differences between the pCR and non-pCR groups regarding pharmacokinetic parameters and tumor volume changes. However, the proportion of voxels with increased signal intensity on the parametric response map demonstrated good diagnostic value for predicting pCR (after the first treatment cycle), with an AUC of 0.770 (95% CI 0.626–0.879) and sensitivity and specificity of 100% and 71%, respectively, at a 21% threshold.

Drisis et al. [55] generated parametric response maps using an affine registration method involving subtraction of pre-treatment images (reference image) from post-NACT images (transformed image). Regions showing an increase in voxel intensity greater than 10% were classified as non-responding (PRM_dce+), whereas regions with a decrease in intensity greater than 10% after treatment initiation were classified as responding (PRM_dce−). The study demonstrated the prognostic potential of both histopathologic characteristics and parametric response maps for predicting non-pCR. The models based on histopathologic features and PRM patterns were shown to achieve AUC values of 0.71 and 0.88, respectively. The authors also found that PRM_dce+ and tumor Grade 2 (moderate differentiation) were significant predictors of non-pCR (AUC = 0.94).

Deep learning methods have also demonstrated promising results in predicting response to NACT [56, 57]. Thus, Comes et al. [56] employed a pretrained convolutional neural network for the automatic extraction of low-level features (edge, line, and point characteristics) from images obtained before and after the first cycle of NACT, thereby eliminating the need for manual segmentation. The study also evaluated the prognostic ability of various molecular features, from which the optimal predictors were selected and subsequently modeled using a support vector machine. The model based solely on molecular features (ER, PR, HER2 status, and molecular tumor subtype) achieved the following performance metrics: accuracy, 69.2%; sensitivity, 42.9%; and specificity, 78.9%. The model integrating both molecular and radiomic features acquired before treatment and at early stages of therapy demonstrated high diagnostic performance on the test dataset: AUC, 0.90; accuracy, 92.3%; sensitivity, 85.7%; and specificity, 94.7%.

Peng et al. [57] compared deep learning and conventional machine learning approaches for predicting treatment response based on baseline molecular, kinetic, and radiomic features. Traditional machine learning methods included manual radiomic feature extraction and application of the least absolute shrinkage and selection operator for optimal feature selection; moreover, linear discriminant analysis was used as a robust supervised classifier. Among deep learning approaches, the ResNeXt50 deep residual neural network was used for radiomic feature extraction, whereas a multilayer perceptron was applied for model construction based on kinetic and molecular features. The AUC values for models using only radiomic, kinetic, or molecular features, developed with linear discriminant analysis and multilayer perceptron, did not exceed 0.75. A slight improvement in performance was observed when these features were combined. However, the convolutional neural network-based model incorporating all features significantly outperformed the linear discriminant analysis model in diagnostic efficiency, achieving an AUC of 0.832 (95% CI 0.816–0.847) and an accuracy of 0.772 (95% CI 0.724–0.821).

Li et al. [58] developed a nomogram based on dynamic contrast-enhanced MRI data to predict pCR in patients with triple-negative breast cancer. Univariate and multivariate logistic regression analyses were used to identify independent predictors of pCR. The nomogram was constructed using three key predictors:

• androgen receptor (AR) status
• tumor volume
• time to peak (TTP).

The resulting nomogram demonstrated high predictive performance, with an AUC of 0.79 in the validation cohort. The study also emphasized that tumors characterized by a TTP of 2 minutes, large tumor volume, and AR positivity were associated with a lower probability of achieving pCR.

CHALLENGES AND LIMITATIONS OF USING RADIOMICS TO PREDICT THE EFFECTIVENESS OF NEOADJUVANT CHEMOTHERAPY

A review of the published data demonstrates that models for predicting response to NACT have evolved substantially with the introduction of machine learning and deep learning methods. Most studies have identified significant correlations between radiomic features and response to NACT. In addition, studies have shown an expanded region of interest to include both the tumor and the adjacent tissues (the peritumoral region) [34, 53], as well as the integration of radiomic features with clinical and molecular characteristics, which opens new perspectives for a more in-depth understanding of tumor biology in the context of NACT [33, 34, 49, 50, 51].

Despite the promising results of predictive radiomic models, there are substantial limitations that hinder clinical implementation of radiomics. Most studies are single-center, retrospective, and involve relatively small patient cohorts [32–34, 49, 55, 57, 58].

Some reports demonstrate contradictory results when using identical predictive features. In addition, the use of manual segmentation for defining regions of interest may introduce variability both within and between observers [18, 19]. Data quality also depends on the diagnostic equipment used and operator expertise, whereas the large number of extracted features complicates analysis and interpretation and requires considerable computational resources. Differences in treatment regimens, molecular tumor subtypes, and the lack of standardization of scanning, analysis, and data-processing protocols contribute to heterogeneity across studies, which directly affects their reproducibility [12, 13, 17, 19, 20].

Future studies should focus on addressing the existing gaps, with priority given to the development of standardized data acquisition protocols that provide clear guidelines for imaging parameters and radiomic feature extraction methods. Their harmonization may reduce variability and enable the achievement of more comparable and reproducible results, thereby strengthening the reliability and scientific validity of predictive models. This process should be based on prospective clinical studies as well as the creation of large, balanced multicenter datasets. In this regard, there is a clear need to develop standards and structured protocols that enable their universal implementation across different centers, users, and imaging equipment [12, 13, 17, 19, 20].

The collection of external and independent validation datasets represents a critically important component for testing predictive models. Radiomic datasets used for training, testing, and validation of artificial intelligence models should be continuously updated and expanded, incorporating comprehensive statistical and clinical parameters to support model integration and performance assessment in clinical practice [12, 13, 17, 19, 20].

The next stage in the development of radiomics involves overcoming the above-mentioned limitations through the use of simpler and more widely available imaging modalities, such as mammography. Mammography remains the most accessible and cost-effective method for breast cancer screening and is widely implemented in most medical institutions. Mammographic examinations are characterized by a high degree of standardization, rapid acquisition, and ease of performance. Focusing on mammography and enhancing its diagnostic value through radiomics and artificial intelligence technologies may substantially increase the predictive performance for response to NACT, providing a more personalized and accessible treatment approach for a larger number of patients [24, 26].

CONCLUSION

Breast cancer remains a global challenge that requires the development of innovative approaches to treatment and prediction of therapeutic efficacy. Quantitative assessment of medical imaging has been recognized by leading professional communities and represents an important step toward personalized medicine. Its mandatory application before NACT may improve treatment outcomes and patient prognosis. Early prediction of NACT effectiveness makes it possible to select the optimal management strategy, thereby reducing unnecessary toxicity, financial costs, and treatment-related adverse effects. However, radiomics still requires time to play a significant role in translational cancer research, and even more time is needed for its widespread implementation into clinical practice.

ADDITIONAL INFORMATION

Supplement 1: Application of conventional and contrast-enhanced spectral mammography for evaluating the response to neoadjuvant chemotherapy in breast cancer patients. doi: 10.17816/DD634972-4348463

Supplement 2: Application of ultrasound for evaluating the response to neoadjuvant chemotherapy in breast cancer patients. doi: 10.17816/DD634972-4348470

Supplement 3: Application of magnetic resonance imaging with dynamic contrast enhancement for evaluating the response to neoadjuvant chemotherapy in breast cancer patients. doi: 10.17816/DD634972-4348474

Author contributions: M.M. Suleymanova: data curation, writing—original draft, writing—review & editing; E.V. Kondratyev: conceptualization; V.A. Nechaev, M.V. Ermoshchenkova: writing—review & editing, supervision; A.Yu. Popov, E.S. Kuzmina: writing—review & editing, supervision; G.G. Karmazanovsky : conceptualization, supervision. All the authors approved the version of the manuscript to be published and agreed to be accountable for all aspects of the work, ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved.

Ethics approval: Not applicable.

Funding sources: No funding.

Disclosure of interests: The authors have no relationships, activities, or interests for the last three years related to for-profit or not-for-profit third parties whose interests may be affected by the content of the article.

Statement of originality: No previously published material (text, images, or data) was used in this work.

Data availability statement: The editorial policy regarding data sharing does not apply to this work.

Generative AI: No generative artificial intelligence technologies were used to prepare this article.

Provenance and peer-review: This paper was submitted unsolicited and reviewed following the standard procedure. The peer review process involved three external reviewers and the in-house science editor.

About the authors

Maria M. Suleymanova

A.V. Vishnevsky National Medical Research Center of Surgery; Moscow City Hospital named after S.S. Yudin

Author for correspondence.
Email: maria.suleymanova95@gmail.com
ORCID iD: 0000-0002-5776-2693
SPIN-code: 7193-6122

MD

Russian Federation, Moscow; Moscow

Grigory G. Karmazanovsky

A.V. Vishnevsky National Medical Research Center of Surgery; The Russian National Research Medical University named after N.I. Pirogov

Email: karmazanovsky@yandex.ru
ORCID iD: 0000-0002-9357-0998
SPIN-code: 5964-2369

MD, Dr. Sci. (Medicine), Professor, academician of the Russian Academy of Sciences

Russian Federation, Moscow; Moscow

Evgeny V. Kondratyev

A.V. Vishnevsky National Medical Research Center of Surgery

Email: evgenykondratiev@gmail.com
ORCID iD: 0000-0001-7070-3391
SPIN-code: 2702-6526

MD, Cand. Sci. (Medicine)

Russian Federation, Moscow

Anatoly Yu. Popov

A.V. Vishnevsky National Medical Research Center of Surgery

Email: vishnevskogo@ixv.ru
ORCID iD: 0000-0001-6267-8237
SPIN-code: 6197-2060

MD, Cand. Sci. (Medicine)

Russian Federation, Moscow

Valentin A. Nechaev

Moscow City Hospital named after S.S. Yudin

Email: dfkz2005@gmail.com
ORCID iD: 0000-0002-6716-5593
SPIN-code: 2527-0130

MD, Cand. Sci. (Medicine)

Russian Federation, Moscow

Maria V. Ermoshchenkova

Moscow City Hospital named after S.S. Yudin; Sechenov First Moscow State Medical University (Sechenov University)

Email: ermoshchenkova_m_v@staff.sechenov.ru
ORCID iD: 0000-0002-4178-9592
SPIN-code: 2557-7700

MD, Dr. Sci. (Medicine)

Russian Federation, Moscow; Moscow

Evgeniya S. Kuzmina

Moscow City Hospital named after S.S. Yudin

Email: saparts@mail.ru
ORCID iD: 0009-0007-2856-5176
SPIN-code: 9668-5733

MD

Russian Federation, Moscow

References

Supplementary files