Technical Summary: Multimodal Fusion for Preoperative Prediction of Lymphovascular Invasion in Breast Cancer

Problem Statement

Lymphovascular invasion (LVI) is a critical prognostic factor in breast cancer, associated with increased risks of local recurrence, distant metastasis, and reduced overall survival. However, preoperative assessment of LVI remains challenging. The current gold standard, postoperative histopathological examination, cannot guide preoperative surgical planning or neoadjuvant therapy selection. While core needle biopsies offer a minimally invasive alternative, they often fail to capture the spatial heterogeneity of the tumor and the peritumoral regions where LVI frequently occurs, leading to sampling errors and significant inter-observer variability. Consequently, there is an urgent need for non-invasive, reproducible imaging methods capable of accurately predicting LVI status prior to surgery.

Methodology

This study employed a retrospective, multicenter design involving 750 patients with pathologically confirmed breast cancer from two independent clinical centers. The data were partitioned into a development cohort (Center 1, n=636), which was further split into a training set (n=445) and an internal validation set (n=191), and an independent external test set (Center 2, n=114).

Data Acquisition and Preprocessing

• Imaging: Dynamic contrast-enhanced magnetic resonance imaging (DCE-MRI) was acquired using 3.0T scanners (Philips and GE) with varying protocols.
• Segmentation: Two experienced radiologists manually delineated tumor regions of interest (ROIs) on the slice showing the largest tumor area, avoiding necrosis and cystic changes.
• Clinical Data: Pathological and clinical variables, including age, menopausal status, histological grade, sentinel lymph node status, immunohistochemical markers (ER, PR, HER-2, Ki-67), and molecular subtypes, were extracted.

Feature Extraction and Model Construction

The study systematically compared four ROI cropping strategies for 2D deep learning input:

1. crop: A rectangular region containing the tumor.
2. crop_roi_only: The tumor ROI only.
3. crop_en5: The ROI expanded by 5 mm into the peritumoral region.
4. no_crop: The entire image slice containing the tumor.

Five predictive models were constructed and compared:

1. Clinical Model (Clinic): Based on independent risk factors identified via multivariate logistic regression.
2. Radiomics Model (Rad): Extracted using PyRadiomics (first-order, shape, texture features) with LASSO regression for selection and five machine learning classifiers (LR, SVM, KNN, RF, MLP).
3. Deep Learning Model (DTL): Features extracted using a VGG19 network (pre-trained on ImageNet) with the optimal ROI strategy, fed into the same five classifiers.
4. Deep Learning Radiomics Model (DLR): A feature-level fusion of traditional radiomics and deep learning features.
5. Combined Model: A comprehensive fusion of radiomics, deep learning features, and clinical risk factors.

Performance was evaluated using AUC, accuracy, sensitivity, specificity, calibration curves, and Decision Curve Analysis (DCA).

Key Contributions

1. Optimization of ROI Strategies: The study provides a systematic evaluation of ROI definition strategies for 2D deep learning in LVI prediction. It identifies that a simple rectangular crop containing the tumor and limited background (crop) yields the best generalization performance, outperforming precise segmentation, peritumoral expansion, and full-slice inputs.
2. Multimodal Fusion Framework: The authors demonstrate the incremental value of integrating traditional radiomics, 2D deep learning features, and clinical risk factors. The study highlights that while standalone deep learning models may struggle with subtle histopathological features like LVI, their integration with radiomics and clinical data significantly enhances predictive capability.
3. Rigorous Multicenter Validation: Unlike many prior studies relying on single-center data, this work employs an independent external test set from a different institution with different imaging protocols, providing a robust assessment of model generalizability.

Results

• ROI Strategy Comparison: The crop strategy demonstrated superior generalization on the external test set, with AUC values ranging from 0.652 to 0.683 across classifiers. In contrast, the no_crop strategy performed the worst (AUC < 0.531), and the crop_en5 strategy showed signs of overfitting during training.
• Clinical Predictors: Multivariate analysis identified menopausal status, histological grade, sentinel lymph node status, Ki-67 expression level, and molecular subtype as independent predictors of LVI.
• Model Performance:
  • The Combined Model achieved the best overall performance on the external test set with an AUC of 0.720 (95% CI: 0.6245–0.8146).
  • Notably, while the Clinical-only model had a slightly higher AUC (0.743), the Combined model exhibited significantly higher sensitivity (0.847 vs. 0.610), reducing the rate of false negatives.
  • Standalone models (Rad, DTL, DLR) showed lower performance on the external test set (AUCs < 0.650), with the DTL model performing particularly poorly (AUC 0.469).
• Clinical Utility: Decision Curve Analysis confirmed that the Combined model provided the highest net benefit across a wide range of threshold probabilities, supporting its clinical utility for risk stratification.

Significance and Claims

The paper claims that the proposed multimodal fusion model offers a promising non-invasive approach for preoperative LVI risk stratification. By achieving high sensitivity, the model addresses a critical clinical need to minimize false negatives, thereby preventing the underestimation of patient risk and inadequate treatment planning. The study emphasizes that the rectangular ROI cropping strategy is the optimal input configuration for 2D DCE-MRI deep learning in this context.

The authors maintain a modest tone regarding their findings, acknowledging limitations such as the retrospective nature of the study, the relatively small external validation set, and the use of 2D single-slice inputs which may not fully capture volumetric tumor biology. They conclude that while the model shows potential for personalized treatment decisions, it should be viewed as a complementary tool to existing clinical assessments rather than a standalone replacement, and further prospective validation is required.