Pixels to Prognosis: Translating Synthetic MRI Radiomics into Clinical Practice for Human Epidermal Growth Factor Receptor 2–Positive Breast Cancer

See also article by Chen and Chen et al in this issue.

Chunhong Hu, PhD, is a professor of medical imaging at The School of Biomedical Engineering and Technology, Tianjin Medical University, Tianjin, China. His research interests focus on medical image processing, with particular emphasis on imaging microstructures in hepatic fibrosis and its regulating role in portal hypertension. He is also dedicated to integrating advanced x-ray phase-contrast CT with deep learning techniques to enhance the quality of image reconstruction and boost x-ray imaging safety performance. He has published more than 90 peer-reviewed articles on medical imaging.

In 2012, Philippe Lambin and colleagues (1) helped reshape the landscape of medical imaging by formally introducing the quantitative imaging framework known as radiomics, enabling the efficient extraction of mineable data from conventional medical images. During its early development, radiomics garnered widespread interest by establishing the concept that comprehensive information regarding the macroscopic tumor phenotype could be obtained noninvasively, without the need for tissue sampling (1,2). Concurrently, quantitative imaging technologies and related approaches advanced the field of personalized oncology, demonstrating broad utility in tumor characterization and treatment response prediction (2).

Radiomics derives quantitative morphologic, intensity, and texture descriptors from standard clinical images. In the mid-2010s, researchers demonstrated the potential value of intratumoral and peritumoral radiomics for predicting pathologic complete response (pCR) following neoadjuvant chemotherapy in breast cancer (3). Over the subsequent decade, numerous research groups worldwide refined radiomic workflows and validated their feasibility across various molecular subtypes. These studies reflected technical advances and suggested that imaging biomarkers might surpass conventional clinical assessment methods; however, their ability to capture the complex morphologic features and vascular architecture of highly aggressive tumors remained limited (3).

Recently, Chen and Chen and colleagues (4) conducted a multicohort study, published in this issue of Radiology: Imaging Cancer, using a synthetic radiomics (S-Radiomics) model to predict pCR in patients with human epidermal growth factor receptor 2–positive breast cancer following neoadjuvant chemotherapy by analyzing pretreatment dynamic contrast-enhanced MRI data. The proposed model integrated data from 404 female patients, including a primary cohort of 283 patients from three independent clinical centers and two independent public validation cohorts (Investigation of Serial Studies to Predict Your Therapeutic Response with Imaging and Molecular Analysis 2 [I-SPY 2] for genomic validation and Investigation of Serial Studies to Predict Your Therapeutic Response with Imaging and Molecular Analysis 1 [I-SPY 1] for external testing). Unlike conventional radiomic approaches, the study combined traditional radiomic features with topological and Hessian-based features and validated the methodology using an internal testing set, an external testing set, and two independent public validation cohorts (4). The inclusion of multiple independent validation cohorts strengthens the robustness and generalizability of the findings for an imaging biomarker study.

Human epidermal growth factor receptor 2–positive breast cancer accounts for a substantial proportion of aggressive breast malignancies. However, predicting which patients will respond to standard targeted neoadjuvant regimens remains a clinical challenge (5). Early identification of pCR is particularly important because achieving this end point is strongly associated with improved long-term outcomes (6). In this study (4), Chen and Chen et al investigated morphologic and vascular structural differences between treatment responders and nonresponders using a synthetic feature extraction approach. By incorporating topological and Hessian-based mathematical features, the proposed model may address some of the limitations of conventional radiomics. Characterizing the three-dimensional geometric structure of the tumor enables differentiation between spatial compactness and complex internal architecture, providing additional insight into tumor organization and vascular heterogeneity.

Topological radiomics enables quantitative characterization of subtle three-dimensional tissue patterns that are difficult to appreciate visually. The proposed S-Radiomics model incorporates multiscale statistical analysis to capture fine spatial gradients and tubular structural characteristics through Hessian-based features (4).

Advanced radiomic feature extraction is typically performed in specialized research environments, and computation of complex topological indexes often requires substantial computational resources. In addition, the lack of standardized and widely accessible tools for multiscale geometric analysis remains a barrier to broader implementation. Nevertheless, this rapidly evolving field continues to show considerable promise for translating advanced radiomics into routine clinical practice (5).

This study (4) differs from previous breast MRI radiomics studies in several important respects. First, Chen and Chen and colleagues evaluated not only the predictive performance of the algorithms but also their underlying biologic mechanisms. Although most patients with human epidermal growth factor receptor 2–positive tumors receive standardized targeted therapies, intrinsic molecular heterogeneity influences therapeutic response (5). The investigators (4) compared the predictive performance of the S-Radiomics model with that of both traditional radiomics and clinical models and further validated the findings using transcriptomic and immunologic profiles.

The S-Radiomics model achieved area under the receiver operating characteristic curve values of 0.86 and 0.84 in the internal and external test sets, respectively, representing a notable improvement over the traditional radiomics model (areas under the receiver operating characteristic curve of 0.78 and 0.73, respectively) (4). Furthermore, significant improvements in the integrated discrimination improvement and net reclassification index, particularly in the training and external test sets, supported the incremental value of the S-Radiomics model. Image quality and standardized processing protocols enabled model interpretation using the Shapley additive explanations method. Shapley additive explanations analysis identified the topological feature Gaussian_curvature_mean as the most influential predictor in the S-Radiomics model. Higher mean Gaussian curvature, reflecting a more convex and compact tumor morphology, was positively associated with pCR, whereas greater shape flatness was negatively associated with achieving pCR.

The investigators also analyzed paired messenger RNA expression data to identify biologic pathways associated with S-Radiomics predictions (4). Gene set enrichment analysis identified pathways associated with the underlying tumor biology. The predicted pCR group demonstrated upregulation of RNA degradation and protein synthesis pathways, along with significantly greater immune cell infiltration, including T lymphocytes and CD8+ T cells.

Numerous international centers now combine advanced MRI platforms with genomic sequencing to perform radiogenomic analyses (5). However, relatively few institutions currently have the capability to routinely integrate topological algorithms into clinical workflows. An important remaining question, therefore, is when this technology can be translated into routine clinical practice.

This question is increasingly being addressed through feasibility studies incorporating technological advances such as machine learning, deep learning, and automated tumor segmentation. Deep learning–based feature extraction has the potential to generate high-quality predictive datasets without manual tumor delineation (7). These approaches may enable predictive performance comparable to that of complex mathematical modeling while substantially reducing manual effort (7). Automated, robust tumor segmentation for routine MRI examinations will be essential for clinical implementation. Although synthetic radiomic features can be extracted from manually delineated regions of interest, manual segmentation remains labor intensive and is unlikely to be practical for routine clinical use. Consequently, developing reliable automated segmentation methods remains an important priority.

Beyond predicting pCR, evaluating the long-term prognostic performance of these models is also essential. Chen and Chen et al (4) evaluated recurrence-free survival in the primary cohort and demonstrated that patients predicted to achieve pCR by the S-Radiomics model had significantly longer recurrence-free survival (P = .02).

In conclusion, the application of synthetic topological and Hessian-based modeling to human epidermal growth factor receptor 2–positive breast cancer represents an important advance in predictive MRI radiomics. The study demonstrated incremental value over traditional radiomic approaches while also providing biologic insight through radiogenomic analyses. Several limitations should be acknowledged, including the retrospective study design and reliance on manual tumor segmentation. Despite these limitations, the findings support continued efforts toward clinical translation and expand the armamentarium of noninvasive prognostic tools for precision breast cancer diagnosis and treatment (4).

Disclosures of conflicts of interest

Please see ICMJE form(s) for author conflicts of interest. These have been provided as supplemental materials.