# Dynamic graph convolution with comprehensive pruning and GNN classification for precise lymph node metastasis detection

## TL;DR

This paper introduces a new AI framework for detecting lymph node metastases in breast cancer with high accuracy and efficiency.

## Contribution

A novel framework combining dynamic graph convolution and pruning techniques for precise lymph node metastasis detection.

## Key findings

- The proposed framework achieved 98.65% classification accuracy on the CAMELYON17 dataset.
- The model outperforms existing methods in segmentation precision and classification performance.
- The framework effectively handles challenges like low contrast and shape variability in lymph node detection.

## Abstract

Early and accurate detection of lymph node metastases is crucial for improving breast cancer patient outcomes. However, current clinical practices, including CT, PET imaging, and microscopic examination, are time-consuming and prone to errors due to low tissue contrast, varying lymph node sizes, and complex workflows. To address the limitations of existing approaches in lymph node segmentation, feature embedding, and classification, this study proposes a novel framework Graph-Pruned Lymph Node Detection Framework (GPLN-DF) that integrates a Dynamic Graph Convolution (DGC) autoencoder with Node Attribute-wise Attention (NodeAttri-Attention) for accurate lymph node segmentation. This segmentation is further refined using Comprehensive Graph Gradual Pruning (CGP) to reduce unnecessary parameters and computational costs. After segmentation, Hessian-based Locally Linear Embedding (HLLE) is applied for effective feature extraction and dimensionality reduction, preserving the geometric structure of lymph node regions. Finally, a Graph Neural Network (GNN) classifier enhanced with CGP is used to classify the segmented lymph nodes as metastatic or non-metastatic based on the extracted features. This comprehensive framework addresses challenges such as small lymph node size, shape variability, low contrast in medical imaging, and high computational burden. The model was evaluated on the CAMELYON17 dataset, achieving a classification accuracy of 98.65%, surpassing existing models in segmentation precision and classification performance.