🤖 Generated with AI assistance · Does not replace individualized medical advice.
AI Act Transparency Notice
This content was generated by an AI assistant to summarize and contextualize published medical literature. It is intended for informational purposes only and does not constitute medical advice, diagnosis, or treatment. Always consult with a qualified healthcare provider for personal medical decisions.
Multimodal DCE-MRI Stacking Fusion for Luminal vs. Non-Luminal Breast Cancer Classification
Key Facts
| Metric | Value | Context |
|---|---|---|
| External Test AUC | 0.854 | Robust performance on independent cohort PMID: 42710818 |
| Luminal Sensitivity | 0.839 | High ability to correctly identify luminal cases PMID: 42710818 |
| Dominant Modality | DL-ResNet50 | Contributed 48.6% of SHAP attribution PMID: 42710818 |
| Sample Size | 396 patients | 220 training, 56 validation, 120 external test PMID: 42710818 |
Why It Matters / Context
Accurate subtyping of breast cancer is critical for guiding systemic therapy, as luminal (hormone receptor-positive) and non-luminal (often HER2-positive or triple-negative) subtypes have distinct prognostic and therapeutic implications. While histopathology remains the gold standard, non-invasive imaging-based classification can facilitate early treatment planning, particularly in neoadjuvant settings where biopsy results may be pending or when assessing tumor heterogeneity.
Previous attempts at imaging-based classification have often relied on single modalities, such as radiomics or deep learning alone, which may lack robustness across different centers due to scanner variability and acquisition protocol differences. This study addresses that limitation by employing a stacking fusion approach. Stacking is an ensemble technique where the outputs (predicted probabilities) of multiple “base” models are used as features for a “meta-classifier.” By integrating clinical data, whole-tumor radiomics, habitat-based radiomics, and deep learning features, the model aims to capture complementary information that no single modality can provide.
The study’s significance lies not only in its performance metrics but also in its rigorous interpretability analysis. By using SHapley Additive exPlanations (SHAP) and ablation studies, the authors provide insight into how the model makes decisions and which components are truly essential. This transparency is crucial for clinical adoption, as “black box” AI models face significant regulatory and trust barriers in oncology.
Details
The study enrolled 396 patients with pathologically confirmed invasive breast cancer from two centers PMID: 42710818. The cohort was split into a training set (n=220), a validation set (n=56), and an external test set (n=120) to ensure generalizability PMID: 42710818.
Four unimodal base models were trained independently:
1. Clinical-MLP: A multilayer perceptron using clinical features.
2. Radiomics-ExtraTrees: An ensemble tree model using whole-tumor radiomic features.
3. Habitat-ExtraTrees: An ensemble tree model using subregion-based (habitat) radiomic features.
4. DL-ResNet50: A deep learning model using transfer learning on DCE-MRI images.
The predicted probabilities from these four models were fused using a stacking framework. The meta-classifier was selected via grid search from nine candidate model families, ultimately identifying a Support Vector Machine (SVM) as the optimal combiner PMID: 42710818.
Performance Comparison
The stacking model outperformed individual unimodal models. On the external test set, the fusion model achieved an AUC of 0.854 PMID: 42710818. This represented a statistically significant improvement over the Clinical-MLP (P < 0.001) and the DL-ResNet50 model (P = 0.048) PMID: 42710818. The model demonstrated a sensitivity of 0.839 for identifying luminal cases, indicating a strong capacity to detect hormone receptor-positive disease PMID: 42710818.
Interpretability and Ablation Analysis
To understand the contribution of each modality, the authors employed two complementary methods: SHAP attribution and leave-one-modality-out ablation.
- SHAP Analysis: This method quantifies the impact of each feature (or in this case, each base model’s output) on the final prediction. SHAP ranked the DL-ResNet50 model as the dominant contributor, accounting for 48.6% of the total attribution PMID: 42710818. This suggests that the deep learning features derived from the raw MRI images carry the most discriminative information for subtyping.
- Ablation Analysis: This experimental approach involves removing one modality at a time to observe the change in model performance. Removing the DL-ResNet50 component caused the largest decline in AUC (-0.033), confirming its critical role PMID: 42710818. Interestingly, removing the Radiomics-ExtraTrees component resulted in a marginal increase in AUC (+0.003), suggesting that this specific modality may have introduced noise or redundancy in the context of the full ensemble PMID: 42710818.
| Modality | SHAP Attribution (%) | Ablation AUC Change | Interpretation |
|---|---|---|---|
| DL-ResNet50 | 48.6% | -0.033 | Dominant and irreplaceable contributor |
| Radiomics-ExtraTrees | Not specified | +0.003 | Potential redundancy/noise; removal slightly improved performance |
| Clinical-MLP | Not specified | Not specified | Significant improvement over standalone model in fusion |
| Habitat-ExtraTrees | Not specified | Not specified | Part of the multimodal ensemble |
Note: Specific SHAP percentages and ablation changes for Clinical-MLP and Habitat-ExtraTrees were not detailed in the abstract provided, but the overall fusion model significantly outperformed Clinical-MLP alone PMID: 42710818.
Limitations
Critical Considerations for Clinical Translation
1. Retrospective Design: The study is retrospective, which introduces potential selection bias and limits causal inference. Prospective validation is necessary to confirm real-world utility.
2. Sample Size: While multicenter, the total cohort of 396 patients is relatively small for deep learning applications. The external test set (n=120) may not fully capture the variability of broader clinical populations.
3. Scanner Heterogeneity: Although the study is multicenter, the specific MRI scanner models and acquisition protocols are not detailed in the abstract. Variability in DCE-MRI protocols (e.g., temporal resolution, contrast dose) can significantly impact radiomics and deep learning feature stability.
4. Subtype Specificity: The model classifies “luminal” vs. “non-luminal.” It does not distinguish between HER2-positive and triple-negative subtypes within the non-luminal group, which have vastly different treatment paradigms.
5. Ablation Paradox: The finding that removing Radiomics-ExtraTrees slightly improved AUC suggests that not all added modalities contribute positively. This highlights the complexity of ensemble design and the risk of overfitting or feature redundancy.
Clinical Takeaway
Clinicians should view this multimodal DCE-MRI model as a promising non-invasive adjunct for early breast cancer subtyping, but its integration into routine practice requires prospective validation and standardization of MRI acquisition protocols to ensure consistent performance across different healthcare settings.
Full Reference
Li Y, Wang P, Zou X, Zhang Q, Wang W. Multimodal DCE-MRI stacking fusion for luminal versus non-luminal classification in breast cancer: A multicenter study with interpretability and ablation analysis. Magnetic Resonance Imaging. 2026. PMID: 42710818 [DOI: 10.1016/j.mri.2026.110796]