Breast MRI may help estimate genomic recurrence risk in oestrogen receptor-positive, HER2-negative breast cancer by measuring how unevenly tissue patterns are distributed within a tumour. A retrospective analysis published in Insights into Imaging developed and tested an MRI-derived intratumoural heterogeneity score in patients from a Chinese cancer centre and an external cohort from the public Duke breast MRI dataset. The score was combined with selected imaging and clinical features to classify patients above or below the established 21-gene recurrence-score threshold. Spatial variation within tumours improved prediction, although limitations in external specificity, study design and imaging consistency remain before clinical use.
MRI Captures Spatial Differences Within Tumours
The 21-gene assay provides a recurrence score that helps estimate distant recurrence risk and guide decisions about adjuvant chemotherapy in early-stage disease. Its cost, turnaround time and technical requirements created the rationale for assessing an imaging-based alternative that could be derived from routine preoperative MRI.
The analysis included an institutional cohort of more than 450 patients and an external group of 230 patients. All had oestrogen receptor-positive, HER2-negative breast cancer, preoperative breast MRI and results from the 21-gene assay. The institutional cases were divided into training and internal test groups, while the public Duke dataset was used for external testing.
Tumours were automatically outlined on contrast-enhanced MRI using a pre-trained segmentation network. Two radiologists checked the contours without seeing the recurrence-score results and made only minor corrections where needed. The images were then standardised to reduce differences between scanners and imaging protocols.
A clustering method divided each tumour into three spatial subregions based on signal intensity. The resulting score reflected the number, size and fragmentation of connected areas across those regions. Higher values indicated a more fragmented and spatially diverse tumour pattern. Conventional radiomic features were extracted separately, allowing global tumour characteristics and local spatial differences to be considered together. The score was designed to provide a simple quantitative measure of heterogeneity that could complement the other selected features.
Combined Features Improve Risk Prediction
The MRI-derived heterogeneity score was higher in patients classified as high risk by the 21-gene assay across the training, internal test and external test cohorts. Used alone, the score demonstrated an ability to distinguish between the two recurrence-risk groups.
Must Read: MRI Radiomics Predict Breast Cancer Survival
Five prediction models were evaluated using different combinations of clinical information, radiomic features and the heterogeneity score. The model that combined all three components performed best. Its area under the curve reached 0.86 in the internal test cohort and 0.82 in the external cohort. These results were higher than those of the model based only on clinical and radiomic features.
Feature-importance analysis also placed the heterogeneity score first among the variables retained in the combined model in all three cohorts. The model maintained predictive performance in subgroups defined by age, menopausal status, pathological type, lymph node status and tumour size.
Several clinical characteristics were also associated with recurrence-score category. Progesterone receptor-positive tumours were more common in the lower-risk group across all cohorts. Higher-grade disease appeared more often in the high-risk group in the training and external cohorts, while larger invasive tumours were more frequent among high-risk patients in the external cohort. Other assessed clinical factors did not show clear differences between the risk groups.
External Testing Reveals Important Limitations
External validation showed that the combined model retained predictive value in data acquired at a different centre and with several scanner systems. However, the decision threshold selected from the training cohort did not transfer evenly. In the external cohort, sensitivity was high but specificity was low, leading to many false-positive high-risk classifications. This limits the model’s immediate clinical applicability and shows that thresholds may need recalibration for different populations and settings.
The retrospective design also creates a risk of selection bias. Although the external cohort strengthened the assessment, prospective validation is still needed. MRI examinations were acquired on different scanners using non-uniform parameters. Signal normalisation and image resampling were applied to reduce these differences, but technical variation may still affect model stability.
The external dataset did not include diffusion-weighted imaging, so its possible contribution could not be assessed. The approach therefore relied on contrast-enhanced MRI for tumour segmentation, radiomic extraction and spatial heterogeneity assessment.
Automatic segmentation was generally stable and required only limited boundary adjustment. Agreement between automatic and corrected tumour contours was high, while the heterogeneity score remained consistent when calculated from both versions. Even so, the model predicted the 21-gene recurrence-score category rather than patient outcomes directly. Further work is needed to establish how the method performs prospectively and how its thresholds should be adapted for clinical use.
MRI-based measurement of spatial tumour heterogeneity adds information to clinical and radiomic features when estimating the 21-gene recurrence-score category in oestrogen receptor-positive, HER2-negative breast cancer. The combined model performed better than clinical and radiomic assessment alone in both internal and external testing, with the heterogeneity score emerging as its most influential feature. External validation nevertheless exposed a high false-positive rate at the selected threshold. Retrospective design, scanner variation and the absence of diffusion-weighted imaging in the external cohort further limit the findings. Prospective validation and setting-specific recalibration are required before clinical application.
Source: Insights into Imaging
Image Credit: iStock
References:
Chen Y, Shi J, Chen J et al. (2026) MRI-based quantification of intratumoral heterogeneity for predicting recurrence risk in ER+/HER2− breast cancer. Insights Imaging; 17, 177.
