📄 medicine

Preoperative Prediction of Vascular Invasion in Breast Cancer Based on Habitat-Based Radiomics Integrating Intratumoral and Multidimensional Peritumoral Features: A Multicenter DCE-MRI Study

This multicenter study demonstrates that a multimodal fusion model combining DCE-MRI-based habitat radiomics from intratumoral and multidimensional peritumoral regions with clinical factors effectively predicts preoperative lymphatic vessel invasion in breast cancer, offering a validated non-invasive tool for personalized risk stratification.

Original authors: Hongen Li, Qingwen Xiao, Yihui Zeng, Xia Wang, Yan Zhang

Published 2026-09-01

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Hongen Li, Qingwen Xiao, Yihui Zeng, Xia Wang, Yan Zhang

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Breast cancer is not a single, uniform disease; it is a collection of tumors that behave in wildly different ways. Some grow slowly and stay put, while others are aggressive, sending tiny clusters of cells into the body's transport systems to start new colonies elsewhere. One of the most critical signs of this dangerous behavior is lymphatic vessel invasion. This occurs when cancer cells break free from the main tumor and slip into the tiny vessels that carry fluid through the body, acting as a highway for the disease to spread. Knowing whether these cells have entered the vessels before surgery is vital for doctors. It helps them decide how much tissue to remove and whether a patient needs stronger treatments afterward. For decades, the only way to know for sure if this invasion has happened was to wait until after the tumor was removed and examined under a microscope. This delay leaves patients and doctors guessing during the most critical planning phase.

Researchers have long hoped that advanced imaging could provide a clear picture of this hidden danger before the knife ever touches the patient. Magnetic resonance imaging, or MRI, is already a powerful tool for looking at breast tissue, showing details that other scans miss. However, standard ways of looking at these images often treat a tumor as a single, solid object. This approach overlooks the complex internal landscape of the cancer, where different areas might have varying blood flow, cell density, and chemical environments. To solve this, a team of scientists set out to build a new kind of map. They wanted to see if they could use the subtle textures and patterns within an MRI scan to predict whether cancer cells had already begun their journey into the lymphatic system, combining the visual data with simple facts about the patient's health to create a reliable guide for treatment.

The study, conducted across two major hospitals in China, brought together data from 742 women with breast cancer. The researchers focused on a specific type of MRI scan called dynamic contrast-enhanced MRI, which tracks how a dye moves through the tissue over time, revealing the speed and pattern of blood flow. They began by drawing a careful outline around each tumor on the scan. Then, they did something more detailed than usual: they looked not just at the tumor itself, but also at the tissue immediately surrounding it. They expanded their view in steps, checking the tissue one millimeter, three millimeters, five millimeters, seven millimeters, and finally ten millimeters away from the tumor edge. This was done to see if the "neighborhood" around the cancer held clues about its aggression.

To understand the inside of the tumor, the researchers used a method inspired by ecology, treating the tumor like a landscape with different habitats. Instead of seeing one big blob, they used computer algorithms to divide the tumor into smaller, distinct zones based on how the cells behaved. Some zones might be dense with cells, while others might be rich in blood vessels. By analyzing these specific "habitats," they hoped to find the unique signatures of the most dangerous parts of the tumor. They fed all this visual information—both the internal habitats and the surrounding tissue layers—into a suite of computer learning programs. These programs were designed to find patterns that human eyes would miss, testing nine different types of mathematical models to see which one could best predict the presence of lymphatic invasion.

The results showed that looking at the tumor in isolation was not enough. The models that only looked at the tumor itself or at very narrow rings of surrounding tissue struggled to be accurate. However, when the researchers combined the data from the seven-millimeter zone around the tumor with the internal habitat maps, the picture became much clearer. They then added two simple pieces of information that doctors already know: whether the patient had gone through menopause and whether the nearby sentinel lymph nodes (the first nodes to catch cancer) showed signs of trouble. This combination created a powerful new tool. When tested on a completely separate group of patients from a different hospital, this combined model correctly identified the risk of lymphatic invasion in about 71 percent of cases where it was present, while also correctly ruling it out in about 75 percent of cases where it was absent.

The study found that the most effective approach was not to rely on just one type of data. A model using only the patient's clinical history was decent, but adding the detailed imaging data significantly improved the ability to catch the dangerous cases. The researchers noted that the seven-millimeter ring around the tumor seemed to capture the most useful information, likely because it includes the active edge where the tumor interacts with the body, without getting lost in the noise of tissue that is too far away. While the computer models were not perfect, they offered a level of insight that was previously unavailable before surgery. The study suggests that by looking at the tumor as a complex, changing environment rather than a static object, doctors can get a much better sense of the threat it poses.

This work does not claim to have solved the problem entirely. The researchers were careful to note that their model is a tool to assist doctors, not a replacement for the final pathological examination. The study was retrospective, meaning it looked back at data that had already been collected, and the number of patients in the final test group was relatively small. There are also differences in the MRI machines used at the two hospitals, which can affect how the images look. Despite these limitations, the findings are significant because they demonstrate that a non-invasive scan can reveal the hidden behavior of a tumor. By integrating the visual complexity of the tumor's microenvironment with straightforward clinical facts, this approach offers a promising way to stratify risk before a patient ever enters the operating room, potentially leading to more precise and personalized treatment plans for breast cancer.

Technical Summary: Preoperative Prediction of Vascular Invasion in Breast Cancer via Habitat-Based Radiomics

Problem Statement

Lymphatic vessel invasion (LVI) is a critical prognostic indicator in breast cancer, associated with local recurrence, distant metastasis, and reduced overall survival. However, the current gold standard for LVI diagnosis is postoperative pathology, which precludes its use in preoperative decision-making. While core needle biopsies are minimally invasive, they often fail to capture the peritumoral region where LVI is most likely to occur. Existing non-invasive radiomics models often treat tumors as homogeneous entities or rely on fixed-width peritumoral expansions, failing to adequately capture the spatial heterogeneity of the tumor microenvironment (TME) and the complex interactions between intratumoral subregions and the surrounding tissue. Furthermore, many existing models lack rigorous external validation across different imaging devices, limiting their generalizability.

Methodology

This retrospective, multicenter study utilized data from 742 breast cancer patients across two independent centers (Center 1: $n=628$ ; Center 2: $n=114$ ). Center 1 data were split into training ( $n=439$ ) and internal validation ( $n=189$ ) sets, while Center 2 served as an independent external test set.

Data Acquisition and Preprocessing

Imaging: Dynamic contrast-enhanced MRI (DCE-MRI) was performed on 3.0T scanners (Philips Ingenia at Center 1; GE Discovery MR750 at Center 2) with varying acquisition parameters to simulate real-world heterogeneity.
Segmentation: Two experienced radiologists manually delineated tumor regions of interest (ROIs) on phase 3 DCE-MRI images.
Peritumoral Expansion: The original ROI was expanded radially to create peritumoral masks at five incremental distances: 1, 3, 5, 7, and 10 mm.
Habitat Generation: To characterize intratumoral heterogeneity (ITH), voxel-based features (entropy and energy) were extracted using a 3×3×3 sliding window. K-means clustering was applied to segment the tumor into functional subregions ("habitats"). The optimal number of clusters was determined using the Calinski-Harabasz (CH) index, resulting in three distinct subregions.

Feature Extraction and Selection

Radiomics: Features were extracted using PyRadiomics, covering first-order statistics, shape, and texture matrices (GLCM, GLRLM, GLSZM, GLDM, NGTDM).
Robustness: Features were screened for robustness using intraclass correlation coefficients (ICC ≥ 0.85) from test-retest and inter-rater analyses.
Dimensionality Reduction: Highly correlated features (Spearman correlation ≥ 0.90) were removed, followed by Lasso logistic regression to select the final predictive features.

Model Construction

Nine machine learning algorithms (Logistic Regression, Naive Bayes, SVM, KNN, Random Forest, ExtraTrees, XGBoost, LightGBM, AdaBoost) were evaluated. The study constructed and compared the following signatures:

Intratumoral (Intra): Features from the tumor core.
Peritumoral (Peri1mm–Peri10mm): Features from specific expansion distances.
Habitat (Habitat): Features derived from the clustered functional subregions.
Clinical (Clinic): Independent predictors identified via multivariate analysis (Menopausal status and Sentinel Lymph Node status).
Combined (Fusion): A multimodal model integrating the optimal ExtraTrees habitat features with clinical predictors.

Key Results

Model Performance

Peritumoral Analysis: Among the various expansion distances, the 7-mm peritumoral region combined with the AdaBoost algorithm demonstrated the best generalization in the external test set (AUC = 0.674), though differences between distances were not statistically significant.
Habitat Analysis: The ExtraTrees algorithm showed the best balance of performance and generalization for habitat features (Test AUC = 0.668), outperforming XGBoost which suffered from overfitting (Train AUC = 0.927 vs. Test AUC = 0.650).
Clinical Predictors: Multivariate analysis identified menopausal status (OR = 2.008) and sentinel lymph node status (OR = 0.157) as independent predictors of LVI.
Fusion Model: The multimodal fusion model (Clinical + ExtraTrees habitat features) achieved the most balanced and robust performance:
- External Test Set AUC: 0.740 (95% CI: 0.646–0.834).
- Sensitivity: 71.2% (significantly higher than the clinical-only model's 64.4%).
- Specificity: 74.5%.
- Precision: 72.8%.

Validation and Utility

Calibration: The fusion model demonstrated excellent calibration, with predicted probabilities closely aligning with observed outcomes in the external test set.
Decision Curve Analysis (DCA): The fusion model provided the highest net clinical benefit across a threshold probability range of 0.2–0.6, outperforming single-modality models and the "treat all" or "treat none" strategies.
Statistical Significance: The Delong test confirmed that the fusion model significantly outperformed the clinical-only model in both the training set ( $P < 0.001$ ) and the external test set ( $P < 0.001$ ).

Key Contributions

Systematic Evaluation of Peritumoral Distances: This study is the first to systematically compare five incremental peritumoral expansion distances (1–10 mm) for LVI prediction, providing empirical data on the spatial scale of microenvironmental influence.
Habitat-Based Imaging: The study successfully applied habitat imaging to LVI prediction, demonstrating that clustering tumors into functional subregions captures heterogeneity more effectively than treating the tumor as a single homogeneous entity.
Rigorous Multicenter Validation: Unlike many single-center studies, this work employed a "development at one center, independent testing at another" strategy using different scanner manufacturers (Philips vs. GE) and protocols, enhancing the credibility of the model's generalizability.
Multimodal Integration: The study established a framework that fuses readily available clinical variables with optimal ExtraTrees habitat features, showing that this combination yields superior sensitivity compared to clinical factors alone.

Significance and Claims

The authors claim that this study offers a promising, non-invasive approach for preoperative risk stratification of LVI in breast cancer. By integrating intratumoral heterogeneity, multidimensional peritumoral features, and clinical risk factors, the proposed multimodal framework provides a more robust basis for clinical decision-making than traditional methods.

The study emphasizes that while the model's performance (AUC 0.740) is not yet sufficient to serve as the sole basis for treatment decisions, it holds significant value as a supplementary tool. It is particularly useful for refining risk stratification in cases where biopsy results are inconclusive regarding LVI status. The authors conclude that this framework supports personalized treatment planning, potentially guiding decisions on neoadjuvant chemotherapy intensity or surgical resection extent, and represents a step toward transitioning from "radiomics" to "radiopathomics" by linking imaging phenotypes to underlying biological mechanisms.