A prediction tool built from preoperative breast MRI and clinical information showed similar performance in two internal groups of patients after breast cancer surgery, according to a peer-reviewed study. The model used three features: a tumor larger than 5 centimeters, axillary lymph node metastasis, and an MRI finding called the adjacent vessel sign. The researchers said the results support further testing, but not clinical use yet.
The model’s reported discrimination score, known as the area under the curve or AUC, was 0.809 in the training group and 0.797 in the validation group. After 1,000 bootstrap checks designed to estimate how much the result might be inflated by the sample, the optimism-corrected AUC was 0.798. The reported 95% confidence intervals were wide, especially in validation, ranging from 0.684 to 0.910, which reflects uncertainty around estimates from a relatively small cohort.
What the researchers studied
The investigators reviewed records from 567 people with stage I to III breast cancer who had surgery between January 2010 and December 2023. Patients had to have histologically confirmed invasive cancer, a preoperative MRI within one week before surgery, no treatment before that scan, and complete clinical and follow-up information. The analysis was retrospective, meaning the researchers looked back at existing records.
After exclusions, 395 eligible patients remained. The exclusions included 62 people lost to follow-up, 44 with incomplete clinical data, and 66 who had received neoadjuvant chemotherapy or a core needle biopsy. Sixty-eight patients developed postoperative recurrence and 327 did not. Recurrence included locoregional recurrence and distant metastasis, identified through postoperative clinical examinations and imaging, with suspicious findings confirmed by biopsy when clinically indicated.
Two radiologists independently reviewed the dynamic contrast-enhanced MRI scans while unaware of the patients’ outcomes, resolving disagreements by consensus. The recurrence and non-recurrence groups were matched on age, family history of breast cancer, and type of surgery. This propensity-score matching produced 204 patients: 68 with recurrence and 136 without. The researchers randomly split them into a training group of 142 and a validation group of 62.
Three features carried the signal
In the multivariable analysis, tumor size above 5 centimeters, the adjacent vessel sign, and axillary lymph node metastasis were independent predictors of recurrence. The adjacent vessel sign refers to an enhancing vessel entering or directly contacting the tumor margin. These findings identify patterns linked with recurrence in this dataset; they do not show that the MRI features themselves cause recurrence.
The researchers combined the three predictors in a logistic-regression model, the statistical method used to estimate a recurrence probability from several inputs. In the reported model, the intercept was −2.977, with coefficients of 1.041 for tumor size above 5 centimeters, 1.051 for axillary lymph node metastasis, and 0.995 for the adjacent vessel sign. The resulting score was presented as a nomogram, a chart intended to turn several patient features into a risk estimate.
The model’s calibration, meaning how closely its estimates matched the observed pattern of recurrence in these groups, was reported as satisfactory. Mean absolute error was 0.087 in training and 0.070 in validation. The authors also reported calibration intercepts of 0.00 and −0.138, slopes of 1.00 and 1.100, and Hosmer-Lemeshow P values of 0.117 and 0.098 for training and validation, respectively.
Why the result remains preliminary
After matching, the standardized mean differences for age and family history were 0.043 and 0.050. Surgery type remained imbalanced, with a standardized mean difference of 0.147, so it was adjusted for in the model. Because the matching covered only a limited set of variables and some imbalance remained, residual confounding cannot be excluded.
The sample also placed pressure on the regression analysis. There were 68 recurrence events and 14 candidate variables, giving an events-per-variable ratio of approximately 4.9, below the stated range of 5 to 10. The validation group contained only 62 patients. The researchers used bootstrap correction, but cautioned that near-ideal training calibration estimates may reflect limited sample size and that overfitting cannot be entirely excluded.
The decision-curve analysis found a positive net benefit across threshold probabilities of approximately 10% to 70%, compared with treating everyone or no one. That result describes the model’s performance in an internal analysis; it does not show that using the nomogram to change surveillance or treatment would improve clinical outcomes.
A tool awaiting a bigger test
At the internally selected cutoff, the training threshold was 0.209, with sensitivity of 0.641 and specificity of 0.860. In validation, the cutoff was 0.206, with sensitivity of 0.568 and specificity of 0.889. Because external threshold validation was not reported, the usefulness of these cutoffs in other hospitals remains unknown.
The authors describe the nomogram as internally validated and say it requires testing in independent, multicenter cohorts before clinical implementation. The model was tested in the matched internal split described above, so its performance beyond that setting has not been established. The study reported support from the Medical Science and Technology Research Foundation of Guangdong Province, grant B2025427.
Paper data and sources
Original title: A preoperative nomogram incorporating semantic MRI features for predicting breast cancer recurrence: a propensity score-matched cohort study
Authors: Zhang C, Xiong Y, Zhu C et al.
Journal/Repository: Frontiers in oncology
Status: Peer-reviewed
First online: 2026-10-09
DOI: Not available
Original paper · Full text