Ultrasound‐ MRI Fusion Deep Learning Radiomics Nomogram: Refining Prognostic Stratification for Triple‐Negative Breast Cancer Treated With Neoadjuvant Chemotherapy

We read with great interest the original article by Cheng et al1 published in this journal, which evaluated the prognostic performance of a deep learning radiomics nomogram (DLRN) integrating ultrasound and multi-sequence magnetic resonance imaging (MRI) for triple-negative breast cancer (TNBC) patients undergoing neoadjuvant chemotherapy (NAC).1 This work is highly innovative and clinically meaningful, as it addresses the long-standing limitations of traditional prognostic tools for TNBC—including the low accuracy of clinical-pathological indicators, the invasiveness and high cost of molecular tests, and the insufficient information utilization of single-modality imaging models.1 The authors' achievement of a high concordance index (C-index) of 0.859–0.887 for disease-free survival (DFS) prediction and effective risk stratification for postoperative recurrence provides a novel non-invasive prognostic approach for TNBC, a subtype with the most unfavorable prognosis among breast cancers.1, 2 Cheng et al's1 study has several notable strengths that set it apart from previous research. First, the integration of ultrasound and multi-sequence MRI combines the real-time functional information of ultrasound (e.g., blood flow dynamics) with the high-resolution morphological and soft tissue contrast of MRI, fully capturing the multidimensional features of TNBC's high heterogeneity.1 Second, the combination of handcrafted radiomics features and automatically extracted deep learning features leverages the interpretability of radiomics and the ability of deep learning to capture fine-grained imaging patterns, which is a valuable exploration of multimodal and multi-method integration in oncologic imaging.1, 3 Third, the multicenter study design and rigorous external validation improve the generalizability of the model, and the clear risk stratification threshold (0.545) provides a straightforward reference for clinical preoperative risk assessment.1 Despite these significant contributions, we believe several aspects of this study warrant further discussion and exploration to enhance the model's clinical utility and generalizability. First, the sample size of the validation cohort (n = 31) is relatively small, with a statistical power of only 0.68 for detecting a medium effect.1 Although the authors acknowledged this limitation, larger multi-center prospective cohorts with extended follow-up (exceeding 5 years) are still needed to verify the model's long-term predictive performance and adjust the risk stratification threshold. Second, the model only integrated clinical N stage as a clinical parameter, while previous studies have shown that molecular indicators such as Ki-67 index and BRCA gene mutation status, as well as pathological indicators such as tumor regression grade, are also independent prognostic factors for TNBC.4, 5 The inclusion of these indicators may further improve the model's predictive accuracy for overall survival (OS), whose C-index (0.800–0.811) was slightly lower than that of the combined MRI + US radiomics model.1 Third, the “black box” problem of deep learning remains unresolved: the ResNet152-based deep learning component extracted abstract features, but the biological correlation between these features and TNBC's recurrence and metastasis mechanisms was not elaborated.1, 6 Combining deep learning features with pathological findings to explore their biological significance will help improve the model's interpretability and clinical acceptance. Fourth, although the authors unified the imaging parameters of the validation centers, residual inter-center variability in MRI protocols (e.g., vendor-specific sequences) and ultrasound operation standards may still affect feature reproducibility.1 Establishing a standardized imaging acquisition and post-processing pipeline is a key step for the clinical translation of this model. In addition, the study found that the radiomics model outperformed the deep learning model in OS prediction, and the DLRN model's OS performance was slightly lower than the radiomics model alone.1 This finding highlights that the predictive efficacy of radiomics and deep learning is affected by multiple factors such as sample size, task complexity, and feature characteristics.1, 7 For TNBC prognosis prediction with relatively limited sample size, high-precision handcrafted radiomics features may have more stable performance, while deep learning models may require larger datasets to exert their advantages in capturing complex patterns.8, 9 Future research can explore the optimal integration strategy of radiomics and deep learning features to maximize their complementary advantages. In conclusion, Cheng et al's1 study is a pioneering work in applying multimodal deep learning radiomics to the long-term prognosis prediction of TNBC after NAC, and the developed DLRN model has great potential for clinical application.1 We believe that addressing the above issues through large-sample prospective studies, multi-indicator integration, and standardized clinical translation will further enhance the model's reliability and practical value, providing more accurate individualized prognostic assessment for TNBC patients and guiding clinical decision-making of NAC and postoperative surveillance. We look forward to the authors' follow-up research to validate and optimize this promising prognostic tool. Data sharing is not applicable to this article as no datasets were generated or analyzed during the current study.