LVI-Detection

This repository contains code and models for detecting Lymphovascular Invasion (LVI) in breast cancer tissues using deep learning techniques.

Table of Contents

• Overview
• Framework
• Results
• Acknowledgements
• Reference

Overview

LVI is a crucial feature in breast cancer, linked to a higher risk of metastasis and poorer prognosis. Manual detection is labor-intensive and prone to variability. This project uses deep learning models, specifically Swin-Transformer and GigaPath, to automate LVI detection in whole-slide images of breast cancer tissue. Trained and evaluated on 90 annotated H&E-stained breast cancer slides, the best model achieved an AUC of 97%, a sensitivity of 79%, and an average of 8 false positives per slide, demonstrating the potential of these models to improve diagnostic accuracy and consistency.

Framework

Figure 1: Overview of the Proposed Framework. (a) Training Phase: (I) LVI foci are annotated by two pathologists, with a third resolving disagreements. (II) Swin-Small uses extracted patches for fine-tuning. (III) GigaPath tessellates the WSI and extracts patch embeddings for binary classification. (b) Inference Phase: (1) Swin-Small predicts patch-wise probabilities using tessellation and sliding window, with postprocessing to identify LVI. (2) GigaPath computes patch probabilities in one step, followed by postprocessing to locate LVI.

Results

Figure 2: Left: A WSI with predictions from Swin-Small. Right: The same WSI with predictions from GigaPath. The predicted LVI locations are highlighted with red boxes, while the ground-truth LVI locations are marked with green boxes.

Acknowledgements

This project uses the code and pre-trained weights provided by GigaPath and timm. We have modified the original code to adapt it for our specific task of LVI detection in breast cancer whole-slide images.

We thank the GigaPath and timm library teams for making their code and model weights publicly available.

Reference

Please consider citing the following paper if you find our work useful for your project.

@InProceedings{,
  title = {},
  author = {},
  booktitle = {},
  pages = {},
  year = {2025},
  volume = {},
  series = {},
  month = {},
  publisher = {},
}