Graph adiabatic diffusion neural networks for distribution-shift breast tumor image classification

Breast tumor images show low intra-class similarity and suffer from distribution shift, posing challenges for recognition tasks. While increasing the number of labeled training data is a common strategy to improve performance, the high cost of expert annotation is another challenge. Semi-supervised learning methods, e.g., Graph Neural Networks (GNNs), which smooth features via graph topology, have the potential to reduce the annotation costs for breast tumor datasets while achieving satisfactory classification performance. To address these challenges, we propose Graph Adiabatic Diffusion Neural Networks (GradiNet), which jointly learn discriminative graph structures for discriminative representation and simulate distribution shift environments. Specifically, we model the discriminative graph structure through a graph-learning objective function and demonstrate its effectiveness theoretically and empirically. Furthermore, we design a GNN feature propagation mechanism for the first time by incorporating the Fourier heat diffusion equation with adiabatic boundary conditions. Hence, the mechanism allows the model to adaptively simulate distribution shifts and enhance its generalization ability on both in-distribution (ID) and out-of-distribution (OOD) data. Extensive experiments on public and private breast tumor ultrasound image datasets demonstrate the superiority and effectiveness of our approach, achieving state-of-the-art performance across multiple evaluation metrics.