HistoMRIFuseNet: retrospective confidence-calibrated decision fusion of independent histopathology and DCE-MRI classifiers
Abstract
Histopathology and dynamic contrast-enhanced magnetic resonance imaging (DCE-MRI) provide different information for breast lesion assessment, but public datasets rarely contain both modalities from the same patients. This study evaluates HistoMRIFuseNet as a retrospective decision-level framework for combining independently trained histopathology and MRI predictors without assuming patient-level correspondence. HistoMorphoMamba performs pathology-based classification of BreakHis histopathology images, while DCE-KineMamba performs BI-RADS-derived lesion-risk classification using LA-Breast DCE-MRI data. Their outputs are combined through confidence- and quality-adjusted logit fusion based on evidential uncertainty, prediction entropy, and image-quality measures. On 1,023 held-out label-consistent unpaired analysis pairs, the framework achieved 0.9462 accuracy, 0.9535 F1-score, 0.8898 MCC, 0.9786 AUROC, and 0.9822 AUPRC. In contrast, the strongest test-label-free validation-tuned weighted fusion achieved 0.9404 accuracy, 0.8635 MCC, and 0.9764 AUROC, while random unpaired fusion decreased to 0.7048 accuracy, 0.4052 MCC, and 0.7614 AUROC. Repeated pairing controls showed substantial dependence on pair construction. With identical frozen modality outputs, logistic stacking achieved 0.9410 accuracy, 0.8612 MCC, and 0.9760 AUROC, compared with 0.9462, 0.8898, and 0.9786, respectively, for confidence-calibrated fusion. The corresponding total-model differences were 0.7 million parameters, 0.5 GFLOPs, and 1.2 ms per inference. These results show larger numerical differences in balanced classification and calibration than in overall accuracy or discrimination. The study therefore supports a bounded retrospective analysis of confidence-aware decision combination rather than a same-patient clinical multimodal diagnostic system. Prospective assessment requires paired histopathology–MRI cohorts with a common pathology-confirmed reference standard.
Related
- Diagnostic Accuracy of Multiparametric Ultrasound Versus Cross-Sectional Imaging for Differentiating Benign and Malignant Cervical Lymph Nodes: A Histopathology-Referenced Retrospective Study
- Pre-admission Nutritional Risk as an Independent Predictor of 28-Day Mortality After Hip Fracture: A Retrospective Cohort Study
- Commentary: Incidence and independent risk factors for postoperative delirium in ICU patients after cardiopulmonary bypass cardiac surgery: a retrospective cohort study
- Tin Sensitization in Belgium and the Netherlands: Complementary Findings From Two Independent Retrospective Studies
- Independent linear association between flow-mediated dilation and predicted intermediate-to-high 10-year ASCVD risk in patients with type 2 diabetes: a retrospective study
BibTeX
@article{yadav2026histomrifusenet,
title = {{HistoMRIFuseNet: retrospective confidence-calibrated decision fusion of independent histopathology and DCE-MRI classifiers}},
author = {Himanshu Yadav and Manish Kumar},
journal = {Scientific Reports},
year = {2026},
month = oct,
doi = {10.1038/s41598-026-74514-3},
url = {https://www.nature.com/articles/s41598-026-74514-3},
note = {CC BY-NC-ND 4.0},
}
RIS
TY - JOUR
TI - HistoMRIFuseNet: retrospective confidence-calibrated decision fusion of independent histopathology and DCE-MRI classifiers
AU - Himanshu Yadav
AU - Manish Kumar
PY - 2026
DA - 2026/10/08
JO - Scientific Reports
DO - 10.1038/s41598-026-74514-3
UR - https://www.nature.com/articles/s41598-026-74514-3
AB - Histopathology and dynamic contrast-enhanced magnetic resonance imaging (DCE-MRI) provide different information for breast lesion assessment, but public datasets rarely contain both modalities from the same patients. This study evaluates HistoMRIFuseNet as a retrospective decision-level framework for combining independently trained histopathology and MRI predictors without assuming patient-level correspondence. HistoMorphoMamba performs pathology-based classification of BreakHis histopathology images, while DCE-KineMamba performs BI-RADS-derived lesion-risk classification using LA-Breast DCE-MRI data. Their outputs are combined through confidence- and quality-adjusted logit fusion based on evidential uncertainty, prediction entropy, and image-quality measures. On 1,023 held-out label-consistent unpaired analysis pairs, the framework achieved 0.9462 accuracy, 0.9535 F1-score, 0.8898 MCC, 0.9786 AUROC, and 0.9822 AUPRC. In contrast, the strongest test-label-free validation-tuned weighted fusion achieved 0.9404 accuracy, 0.8635 MCC, and 0.9764 AUROC, while random unpaired fusion decreased to 0.7048 accuracy, 0.4052 MCC, and 0.7614 AUROC. Repeated pairing controls showed substantial dependence on pair construction. With identical frozen modality outputs, logistic stacking achieved 0.9410 accuracy, 0.8612 MCC, and 0.9760 AUROC, compared with 0.9462, 0.8898, and 0.9786, respectively, for confidence-calibrated fusion. The corresponding total-model differences were 0.7 million parameters, 0.5 GFLOPs, and 1.2 ms per inference. These results show larger numerical differences in balanced classification and calibration than in overall accuracy or discrimination. The study therefore supports a bounded retrospective analysis of confidence-aware decision combination rather than a same-patient clinical multimodal diagnostic system. Prospective assessment requires paired histopathology–MRI cohorts with a common pathology-confirmed reference standard.
SN - 2045-2322
ER -