Technical Summary: LV-YOLO for Lymphovascular Invasion Detection in Gastric Cancer

Problem Statement

Lymphovascular invasion (LVI) is a critical prognostic indicator in gastric cancer (GC), serving as a major predictive risk factor for lymph node metastasis. However, routine manual screening for LVI in hematoxylin and eosin (H&E) stained whole slide images (WSIs) faces significant challenges:

• Labor Intensity: LVI lesions are minute and sparsely distributed, requiring pathologists to screen extensive tumor areas under a microscope.
• Subjectivity and Error: Morphological artifacts, such as tissue shrinkage (retraction artifacts) and artificial displacement of tumor cells during sectioning, closely mimic true tumor emboli. This leads to high inter-observer variability and misdiagnosis.
• Resource Constraints: While immunohistochemical staining (e.g., D2-40) can clarify lymphatic boundaries, its high cost and the global shortage of pathologists limit its widespread application.
• Data Scarcity: Despite the prevalence of gastric cancer, deep learning research specifically targeting LVI detection in this context remains limited compared to other malignancies.

Methodology

The authors propose LV-YOLO, a two-stage deep learning framework designed to improve the objectivity and efficiency of LVI detection. The workflow consists of a macroscopic initial screening phase followed by a microscopic precise localization phase.

1. Dataset and Preprocessing

• Data Source: A multicenter dataset comprising 100 gastric cancer WSIs (70 from TCGA-STAD, 30 from Baotou Medical College) was used for training and internal validation. An external dataset of 48 WSIs was used for cross-validation.
• Annotation: Positive LVI regions were annotated by two pathologists and confirmed by a senior pathologist.
• Preprocessing:
  • Stain Normalization: The Vahadane algorithm was applied to mitigate staining heterogeneity.
  • Hard Negative Mining: A progressive strategy was employed. An initial model trained on a 1:1 positive-to-simple-negative ratio was used to identify "hard negative" patches (background areas frequently misclassified as positive). These were injected back into the training set at a 1:2 ratio (positive to hard negative), creating a final 1:3 ratio to reduce false positives.
  • Augmentation: Online dynamic augmentation (flipping, rotation) was applied to positive samples to address class imbalance.

2. Stage One: Macroscopic Screening

• Model: MobileNetV3 was selected as the lightweight classification backbone after evaluating several architectures (ResNet50, VGG16, etc.).
• Function: It performs a global scan of the WSI to generate a risk heatmap, identifying high-risk regions for further analysis.
• Performance: Achieved an AUPRC of 0.9527 and AUROC of 0.9830 at the patch level.

3. Stage Two: Microscopic Localization (LV-YOLO)

The core contribution is the LV-YOLO object detection model, an optimized version of the YOLOv11 architecture. It incorporates two novel modules to address the specific challenges of LVI detection:

• CAFM (Convolution and Attention Fusion Module):

  • Purpose: To perceive the complex tumor microenvironment by integrating local and global features.
  • Structure: A dual-branch architecture.
    • Local Branch: Uses depthwise separable convolutions and channel shuffling to extract fine-grained, cell-level morphological features.
    • Global Branch: Utilizes a self-attention mechanism to capture long-range spatial correlations between tumor emboli and surrounding vascular structures.
  • Benefit: Helps distinguish genuine LVI from artifacts by understanding both cellular details and contextual vascular relationships.

• CARAFE (Content-Aware ReAssembly of FEatures):

  • Purpose: To reconstruct the geometric boundaries of minute lesions during upsampling.
  • Mechanism: Replaces traditional static interpolation with a dynamic, content-aware kernel prediction. It generates reassembly kernels based on local semantics to preserve structural integrity and fine-grained details during feature map expansion.
  • Benefit: Prevents edge blurring of small LVI lesions, which is common in standard upsampling methods.

4. Evaluation Metrics

Performance was assessed using:

• Patch Level: Mean Average Precision (mAP50), AUROC, AUPRC, Accuracy, and F1 Score.
• WSI Level: Sensitivity and Specificity, utilizing a max-pooling multiple instance learning strategy where the slide-level prediction is determined by the highest patch-level confidence.

Key Results

Patch-Level Performance

• LV-YOLO vs. Baselines: The proposed LV-YOLO achieved a mAP of 95.21% and an AUROC of 0.9805, outperforming the baseline YOLOv11 (mAP 93.35%) and other mainstream detectors like Faster R-CNN (mAP 80.37%), RT-DETR (mAP 91.60%), and YOLOv8 (mAP 92.31%).
• Ablation Study:
  • Adding CAFM alone increased mAP to 94.27%.
  • Adding CARAFE alone increased mAP to 94.13%.
  • Combining both modules yielded the highest performance (mAP 95.21%, F1 0.9353), confirming their synergistic effect.
• Attention Mechanism Comparison: CAFM outperformed other attention mechanisms (CBAM, CA, SimAM, etc.) in both mAP and F1 score.

WSI-Level Performance

• The integrated workflow achieved a sensitivity of 0.9226 and specificity of 0.9562 on the internal test set.
• The AUPRC at the WSI level reached 0.9590.

External Validation

• On an independent external dataset with different staining and preparation protocols, LV-YOLO maintained robust performance with an AUPRC of 0.9313 and an F1 score of 0.8527, demonstrating generalization capabilities despite domain shifts.

Failure Analysis

The authors identified specific failure modes:

• False Negatives: Occurred when lymphatic vessel walls were histologically obscure or when tumor cells were tightly integrated with the vessel wall ("mural growth") without clear retraction clefts.
• False Positives: Triggered by artificial clefts created during sectioning or dense clusters of inflammatory cells (lymphocytes) that mimic cancer cell morphology.

Key Contributions

1. Two-Stage Framework: Developed a collaborative workflow combining a lightweight classifier (MobileNetV3) for rapid screening and a specialized detector (LV-YOLO) for precise localization.
2. Novel Architectural Modules: Introduced the CAFM module to fuse local and global features for better microenvironment perception and the CARAFE module to preserve geometric boundaries of minute lesions during upsampling.
3. Hard Negative Mining Strategy: Implemented a progressive mining strategy to construct a challenging training set, effectively reducing false-positive rates in complex tumor microenvironments.
4. Comprehensive Evaluation: Provided rigorous validation across patch-level, WSI-level, and external cross-center datasets, demonstrating stability against staining variations.

Significance and Claims

The paper claims that the proposed LV-YOLO framework provides a cost-effective, objective, and efficient auxiliary screening tool for pathologists. By automating the detection of LVI in routine H&E sections, the system aims to:

• Reduce the labor-intensive burden of manual screening.
• Minimize diagnostic discrepancies caused by morphological artifacts.
• Assist in clinical risk stratification and prognostic evaluation for gastric cancer patients.

The authors emphasize that the system is designed to assist rather than replace pathologists. It functions as a preliminary screening tool to flag high-risk regions for expert review, thereby improving diagnostic efficiency while maintaining the final diagnostic authority with medical professionals. The study acknowledges limitations, including the relatively small sample size (100 WSIs) and the reliance on H&E consensus annotations rather than immunohistochemistry as a gold standard, positioning the tool as a robust aid for preliminary screening rather than a definitive standalone diagnostic.