AI model detects lymphovascular invasion in breast cancer MRI scans

Artificial intelligence may soon be able to detect one of breast cancer’s most dangerous hidden features before a surgeon ever makes an incision. A team of researchers in China has developed a deep learning system that predicts lymphovascular invasion, or LVI, in invasive breast cancer directly from contrast-enhanced MRI scans, according to a study published in BMC Medical Imaging. LVI, the presence of tumor cells within lymphatic or blood vessels surrounding a tumor, is one of the strongest indicators that cancer may spread to lymph nodes or distant organs, yet it can currently only be confirmed by pathologists examining tissue under a microscope after surgery. The new system aims to change that, giving clinicians a reliable, noninvasive estimate of LVI risk at the preoperative stage, when treatment planning decisions about surgery extent, lymph node biopsy, and neoadjuvant therapy are still open.

The research, led by Junyu Lin, Zichang Ma, Yuxi Tao, and colleagues at the Fifth Affiliated Hospital of Sun Yat-sen University in Zhuhai, with corresponding author Yaqin Zhang, was built on a retrospective cohort of 288 patients with pathologically confirmed invasive breast cancer. Every patient had undergone preoperative dynamic contrast-enhanced magnetic resonance imaging, or DCE-MRI, the standard imaging technique that tracks how a gadolinium-based contrast agent flows into and washes out of breast tissue. Because tumors with lymphovascular invasion often show characteristic enhancement kinetics, aggressive contrast uptake followed by rapid washout, DCE-MRI contains subtle information about tumor biology that radiologists can only partially exploit. The team’s goal was to extract that information systematically using artificial intelligence.

The pipeline the researchers designed has two major stages. The first is automated tumor segmentation. Rather than relying solely on radiologists to manually trace tumor boundaries on each MRI slice, a laborious process subject to interobserver variability, the team trained a ResUNet++ network to perform the task automatically. ResUNet++ is an evolution of the widely used U-Net convolutional architecture for medical image segmentation, incorporating residual connections, attention blocks, and nested dense convolutions to improve boundary accuracy and small-structure capture. Its performance was strikingly consistent: the model achieved a Dice coefficient of 0.916 on the internal validation cohort and 0.921 on the external cohort, where the Dice score measures the spatial overlap between the automated segmentation and the manually delineated ground truth, with 1.0 representing perfect agreement. Scores above 0.9 are generally considered excellent for tumor segmentation tasks.

After segmentation, the system applies a boundary dilation of 4 millimeters around the automatically detected tumor region of interest, a deliberate technical choice that captures the peritumoral microenvironment. This matters because lymphovascular invasion occurs in the tissue immediately surrounding the tumor, where tumor cells invade vessel walls, so the peritumoral zone often carries stronger predictive signals than the tumor core itself. Within these dilated regions, the researchers extracted radiomic features, high-dimensional quantitative descriptors of texture, intensity distribution, shape, and spatial heterogeneity that human readers cannot perceive. In parallel, radiologists assessed conventional MRI features according to the BI-RADS criteria, the standardized Breast Imaging Reporting and Data System lexicon, providing a structured human interpretation layer.

The second stage is classification. The team constructed both single-modality models and a multimodal fusion network that integrates three complementary streams of information: the enhanced MRI images themselves, the radiomic features derived from the segmented tumor and its peritumoral region, and the BI-RADS-based semantic features assessed by radiologists. The classifier architecture is transformer-based, meaning it relies on self-attention mechanisms, the same core technology behind modern large language models. Self-attention allows the network to weigh the relationships among all parts of its input simultaneously rather than processing information only through local receptive fields, which is particularly well suited to capturing long-range spatial patterns within and around tumors and to reconciling heterogeneous feature types from different modalities. The final design used a two-stage multimodal classifier in which features are first refined within each modality before being fused.

The performance gains from multimodality were substantial. On the internal cohort of 238 patients, the two-stage multimodal model achieved an area under the receiver operating characteristic curve, or AUC, of 0.873, compared with 0.801 for the best single-modality transformer model. On the independent external cohort of 50 patients from outside the training distribution, the multimodal system scored 0.845 against 0.762 for the best unimodal model. The differences were statistically confirmed using the DeLong test, a standard nonparametric procedure for comparing correlated ROC curves. An AUC above 0.85 in an external, cross-center setting is a notable result for a prediction task of this kind, suggesting the model learned biologically meaningful patterns rather than idiosyncrasies of a single scanner or population.

Just as important as the accuracy is the system’s interpretability. Black-box predictions are a persistent obstacle to clinical adoption of medical AI, so the researchers applied two widely used explanation techniques. Grad-CAM, or Gradient-weighted Class Activation Mapping, generates heatmaps highlighting the image regions most influential in the network’s decision, allowing radiologists to verify that the model attends to tumor and peritumoral areas rather than artifacts. SHAP, which stands for SHapley Additive exPlanations, quantifies the contribution of each radiomic and clinical feature to individual predictions, drawing on game-theoretic Shapley value theory to distribute credit fairly among features. Together, these tools offer transparency into why the model flags a given tumor as likely to exhibit lymphovascular invasion.

The clinical implications could be significant. Currently, the gold standard for LVI assessment is postoperative histopathology using hematoxylin-eosin staining and immunohistochemical markers such as D2-40 and CD31 to visualize lymphatic and vascular endothelium. Because LVI status is only known after surgery, its influence on preoperative decision-making is indirect. Knowing a patient’s LVI risk beforehand could inform whether axillary lymph node dissection is warranted, whether sentinel lymph node biopsy alone is sufficient, whether neoadjuvant systemic therapy should be considered before surgery, and how aggressively to manage breast-conserving approaches. For patients identified as high-risk, clinicians could intensify surveillance and tailor adjuvant therapy planning.

The external validation deserves particular emphasis. Many promising AI models in radiology fail when moved to new hospitals because of differences in scanner manufacturers, imaging protocols, and patient demographics, a phenomenon often described as domain shift. Testing the model on 50 patients from a distinct cohort, while smaller in size, provides early evidence of generalizability across centers. The near-identical segmentation performance between internal and external data, 0.916 versus 0.921 Dice, and only a modest drop in classification AUC from 0.873 to 0.845, suggest the pipeline is reasonably robust to such variation. Still, the external cohort’s size means larger prospective multicenter trials will be needed before the system can be deployed in routine practice.

The study also illustrates a broader trend in medical imaging AI: the move toward fully automated end-to-end pipelines. By chaining automatic segmentation with multimodal transformer classification, the system removes a major bottleneck, the manual delineation of tumor regions, that has limited the scalability of earlier radiomics studies. Manual segmentation typically requires an experienced radiologist to spend twenty to thirty minutes per case, whereas the automated approach can process a scan in seconds, making population-scale screening and analysis feasible. The integration of BI-RADS semantic features alongside deep image features further shows how human expertise and machine perception can be combined, with the transformer architecture acting as a fusion engine that reconciles different levels of abstraction.

Funding for the work came from the National Natural Science Foundation of China, the Basic and Applied Basic Research Foundation of Guangdong Province, and the Zhuhai Basic and Applied Basic Research Project Foundation. The retrospective study was approved by the Medical Ethics Committee of the Fifth Affiliated Hospital of Sun Yat-sen University, with the requirement for individual informed consent waived given the retrospective use of routinely collected clinical data. The article is published open access under a Creative Commons license, and the team has shared supplementary material detailing the methods.

Breast cancer remains the most commonly diagnosed cancer in women worldwide, and lymphovascular invasion is a key element of prognostic staging across all major clinical guidelines, including those of the American Joint Committee on Cancer and the College of American Pathologists. A validated, automated, preoperative LVI predictor would add a genuinely new piece of information to the preoperative decision toolkit, one derived entirely from an imaging examination most breast cancer patients already undergo. If future prospective studies confirm the performance reported here, transformer-based multimodal analysis of DCE-MRI could become a routine companion to the radiologist’s report, quietly flagging the tumors whose behavior is more aggressive than their appearance suggests and helping ensure that surgical and systemic treatment decisions are made with the fullest possible picture of each patient’s disease.