NEED FOR BETTER RESPONSE ASSESSMENT IN BREAST CANCER: NEOADJUVANT THERAPY (NAT) Breast cancer is the most common malignancy among women worldwide, and NAT, including systemic regimens such as neoadjuvant chemotherapy (NACT), has become a standard component of care for locally advanced and biologically high-risk early breast cancer.1 Pathologic complete response (pCR) after NAT is strongly associated with improved long-term outcomes, including event-free and overall survival. Conventional imaging assessment, however, relies primarily on changes in tumor size or Response Evaluation Criteria in Solid Tumors (RECIST) criteria, which insufficiently capture intratumoral heterogeneity and often fail to provide reliable early prediction of treatment response.2,3 Radiomics has emerged as a promising paradigm that extracts high-dimensional quantitative features from routine medical images, such as MRI, ultrasound, mammography, and positron emission tomography-computed tomography (PET/CT) with the extracted data subsequently analyzed using machine learning or deep learning algorithms.1,4 In the context of NAT for breast cancer, quantitative radiomic features are leveraged to predict response, monitor treatment dynamics, evaluate residual disease, and estimate prognosis, thereby enabling more individualized treatment decisions. RADIOMICS WORKFLOW AND NAT-SPECIFIC CONSIDERATIONS A typical radiomics workflow consists of standardized image acquisition and preprocessing, tumor, and peritumoral segmentation, feature extraction (shape, texture, and statistical features), feature selection or dimensionality reduction, and model building with internal and external validation. The Image Biomarker Standardisation Initiative (IBSI) has standardized the processes of image data acquisition, processing and analysis, thus improving reproducibility across different manufacturers and software platforms.4 From a technical standpoint, several design choices are particularly important when building radiomics models for breast cancer NAT. Rather than relying on very large and redundant feature sets, we favor a relatively compact, IBSI-compliant core of first-order shape and texture (a limited number of families, e.g., GLCM, GLRLM, and GLSZM) features, combined with a few carefully selected higher-order or wavelet features. Feature redundancy is typically addressed using correlation-based filtering or clustering.4 For model building in typical single-center or moderate-size cohorts, regularized logistic regression, support vector machines, tree-based ensembles, and gradient boosting are often more appropriate than highly complex deep networks, which we view as better suited for segmentation or very large, multi-institutional datasets.1,2 Robust validation requires nested cross-validation, clear separation between feature selection and testing, and, wherever possible, external validation in temporally or geographically distinct cohorts; otherwise, the reported performance is likely to be optimistic.2 When acquisition protocols differ, feature harmonization or ComBat-like correction and transparent reporting following Transparent Reporting of a multivariable prediction model for Individual Prognosis or Diagnosis (TRIPOD) and the Checklist for Artificial Intelligence in Medical Imaging (CLAIM) are, in our view, more critical than the choice between any two specific machine learning algorithms. In the NAT setting, several aspects are particularly important. First, multi-timepoint imaging (baseline, early-cycle, and preoperative scans) allows modeling of temporal changes in tumor phenotype.5,6 Second, features extracted from the intratumoral region, as well as from peritumoral tissue and background parenchyma, can capture tumor-microenvironment interactions and treatment-induced changes.7,8 Third, integration of radiomic signatures with clinical, pathological, and molecular variables is increasingly recognized as vital for robust predictive models and bridging imaging phenotypes with therapeutic decision-making.1,2,8 The major challenges facing multi-timepoint analysis include ensuring comparability of images pre- and post-treatment, and maintaining segmentation stability amid evolving tumor regression patterns. RADIOMICS FOR SHORT-TERM RESPONSE PREDICTION (SINGLE AND MULTI-TIMEPOINT) Radiomics has been most extensively explored for short-term response prediction, including clinical response and pCR, using both single-timepoint and multi-timepoint imaging. Regarding single-timepoint models, most research has focused on baseline multiparametric MRI. For example, Liu et al. demonstrate with a multicenter cohort that a radiomics model that integrates baseline dynamic contrast-enhanced MRI (DCE-MRI) and diffusion-weighted imaging (DWI) features with clinical variables can predict the likelihood of pCR following NAT in breast cancer, with areas under the curve (AUCs) of > 0.80 for both training and validation sets.8 Gilad et al. used physiologically decomposed diffusion-weighted MRI to derive radiomic signatures, showing that machine learning models based on these features effectively distinguish responders from non-responders.7 Beyond MRI, baseline radiomics from other modalities also contribute to enhancing short-term response prediction. The results of a meta-analysis conducted by Valizadeh et al. indicate that, with regard to predicting response to NAT, ultrasound-based machine learning models achieve promising sensitivity and specificity, substantiating the utility of sonographic radiomics at baseline.3 Radiomic parenchymal phenotypes extracted from mammographic texture have been linked to breast cancer risk and underlying tissue biology, and may provide additional context for interpreting NAT benefit at the level of the individual.9 Multi-timepoint radiomics is used to capture the dynamic evolution of the tumor phenotype under treatment. Xu et al. propose a fusion model that integrates multi-sequence MRI radiomics with “habitat” imaging features across several NAT cycles and report that this longitudinal model outperforms single-timepoint approaches in predicting pCR.10 Quantitative ultrasound (QUS) radiomics has also been used longitudinally-in a randomized feasibility study. Dasgupta et al. showed that early-treatment changes in QUS-derived parametric features can predict subsequent pathological responses and be used to guide adaptive NAT strategies.5 Furthermore, ultrasound-based nomogram models have been developed to classify pCR versus non-pCR using pre- and intra-treatment sonographic characteristics, further underscoring the role of serial ultrasound radiomics in early response assessment.11 Taken together, these research findings indicate that both single-timepoint and multi-timepoint radiomic models across MRI, ultrasound, and QUS can provide clinically relevant early response information beyond conventional size-based criteria, and may help identify patients unlikely to benefit from standard NAT regimens at an early stage.3,5,9,11-13 In our view, a pragmatic strategy for clinical translation is to combine a robust baseline MRI model with very early ultrasound or QUS changes, which keeps the pipeline relatively simple while still capturing both the intrinsic tumor phenotype and the treatment-induced dynamics. We also consider multi-timepoint models most useful when they are explicitly designed around concrete clinical decisions, such as whether to continue, switch, or intensify NAT, rather than as purely academic exercises in longitudinal modeling. RADIOMICS FOR PREDICTION OF RESIDUAL CANCER BURDEN Residual cancer burden (RCB) after NAT describes how much tumor tissue remains and is widely used to guide postoperative management. In current clinical practice, RCB is determined using surgical specimens, which limits preoperative planning. Radiomics provides an opportunity to approximate RCB noninvasively by quantifying residual disease on imaging at the end of therapy. Multiparametric MRI radiomics can capture the volume, morphology, and internal heterogeneity of residual enhancing tissue and peritumoral changes, offering a more detailed description of residual burden than simple diameter measurements.14,15 Ultrasound-based and QUS-based models have similarly been used to distinguish a minimal or no residual lesion from a more extensive remaining tumor, providing an imaging-based stratification of residual burden that parallels low versus high RCB categories in clinical practice. These reports indicate that multi-modal radiomics (MRI, ultrasound, and QUS) can substantially enrich imaging assessment of residual tumor amount and distribution, and may ultimately complement pathological RCB in tailoring surgery and adjuvant systemic therapy.5 RADIOMICS AND RADIOGENOMICS FOR LONG-TERM OUTCOME PREDICTION Beyond short-term response and residual disease, radiomics has been explored for predicting long-term outcomes, including recurrence risk and survival. Li et al. developed MRI-based radiomic signatures that predict recurrence risk categories as defined by research versions of MammaPrint, Oncotype DX, and PAM50 gene assays, indicating that imaging features can act as noninvasive surrogates for expensive and invasive genomic tests.16 In the NAT context, QUS radiomics has been shown to not only be linked to immediate response but to also predict recurrence in patients with locally advanced breast cancer, demonstrating that ultrasound-derived microstructural features contain prognostic information beyond conventional imaging and clinicopathologic factors.17 Radiogenomics further combines radiomic features with transcriptomic or other multi-omics data to explain why certain imaging phenotypes are associated with outcomes. In a study involving triple-negative breast cancer patients who received NAT, Zhou et al. built a preoperative radiogenomics model that integrates quantitative MRI heterogeneity metrics, gene-expression data, and clinicopathologic variables; this model can predict both pCR and long-term outcomes, thereby linking imaging features to treatment sensitivity and prognosis.13 Zhang et al. provide a comprehensive review of spatial-temporal radiogenomics strategies in breast cancer, highlighting how longitudinal imaging combined with multi-omics data can be used to capture evolving tumor biology during and after NAT, and to improve the prediction of NAT efficacy and long-term prognosis.12 Collectively, these studies support the postulate that radiomics and radiogenomics-based on single or multiple timepoints-may yield noninvasive, biologically informed biomarkers for long-term risk stratification and individualized follow-up strategies in breast cancer patients treated with NAT. ADVANTAGES, CHALLENGES, AND FUTURE DIRECTIONS FOR CLINICAL TRANSLATION In the NAT setting, radiomics offers several advantages: it is noninvasive and repeatable; it captures the entire tumor and its microenvironment rather than just a small biopsy sample; and it can be naturally embedded into routine MRI, ultrasound, and mammography workflows. Existing models have shown promise for early response prediction, individualized treatment adaptation, refined assessment of residual disease, and long-term risk stratification. However, clinical translation remains challenging. Most studies are single-center, retrospective, and small, with limited external validation, leading to concerns about overfitting and poor generalizability. Substantial heterogeneity in imaging protocols, scanner vendors, reconstruction parameters, and segmentation strategies undermines feature stability and model robustness, underscoring the need for IBSI-compliant, harmonized pipelines. Many deep learning models still behave like “black boxes”, with limited interpretability and weak biological plausibility, which reduces clinician trust and complicates regulatory approval. Beyond these methodological issues, there are also very practical barriers to integration. In most hospitals, images are stored and viewed in vendor-specific picture archiving and communication systems (PACS) and diagnostic workstations, whereas radiomics tools typically run as separate research software. Making these systems communicate requires interoperable interfaces: DICOM-based image exchange, vendor-neutral archives, and, where available, HL7/FHIR or vendor APIs for transferring metadata and structured results. We envisage radiomics being deployed as lightweight services that receive research data from PACS, run inference, and then generate and export concise, structured outputs that can be reviewed in the existing reporting environment and discussed at multidisciplinary team meetings. For routine clinical use, inference speed and workflow fit are as important as accuracy. Deep learning-based segmentation and feature extraction should ideally be completed within a few seconds per case, and we therefore favor streamlined models, efficient implementations, and, where possible, on-premise graphics processing units (GPUs) or edge-computing resources rather than latency-prone remote servers. Although fully automatic segmentation is increasingly feasible, we see AI-assisted, semi-automatic contours with radiologist editing as a pragmatic compromise that preserves clinical oversight while still reducing the manual workload. Standardized reporting according to IBSI, TRIPOD, and CLAIM will also be important for regulatory review and for comparing systems across centers. Ultimately, in our view, successful real-world implementation will depend less on marginal gains in AUC and more on fast and reliable inference, robust and auditable segmentation, seamless integration with PACS and multidisciplinary team (MDT) workflows, and solutions that fit within regulatory, infrastructural, and training constraints. If these conditions are addressed, radiomics and radiogenomics have a realistic chance of contributing to transitioning NAT from an experience-based approach to a data-driven precision therapy. Acknowledgement None. Source of funding This work was supported by the National Natural Science Foundation of China (82373231, 82403132), the Natural Science Foundation of Liaoning Province of China (2024-BS-060), and Chinese Young Breast Experts Research Project (CYBER-2021-A02, CYBER-2022-001). Author contributions Wang YS and Xu N: Writing—original draft. Xu YY and Wang MZ: Writing—review & editing. All authors have read and approved the final manuscript. Informed consent Not applicable. Ethical approval Not applicable. Conflict of interest The authors declare that they have no conflicts of interest. Use of large language models, AI and machine learning tools None declared. Data availability statement No additional data.
Wang et al. (Fri,) studied this question.