DGTN: Graph-Enhanced Transformer with Diffusive Attention and Gating Mechanism for Multi-Task Breast Ultrasound Tumor Analysis.
Authors
Affiliations (5)
Affiliations (5)
- Department of Computer Control and Management Engineering, Antonio Ruberti, Sapienza University of Rome, 00185 Roma, Italy.
- Institute of Computer Science, Shah Abdul Latif University, Khairpur 66020, Pakistan.
- Division of Oncology, Washington University, St. Louis, MO 63130, USA.
- Department of AI and Robotics, Sejong University, Seoul 05006, Republic of Korea.
- Computer Vision Laboratory, CAIDAS & IFI, University of Wurzburg, 97070 Wurzburg, Germany.
Abstract
Early and accurate analysis of breast cancer is critical for improving patient outcomes. Ultrasound imaging is widely used for breast tumor screening due to its safety, accessibility, and low cost. We propose DGTN (Diffused Graph-Transformer Network), a lightweight, end-to-end deep learning model that jointly performs breast tumor segmentation and multi-class classification (benign, malignant, and normal) from ultrasound images. DGTN integrates Graph Convolutional Networks (GCNs) and Transformer encoders through a bidirectional diffusive attention mechanism and a learnable gating strategy, enabling structured spatial information and global contextual features to co-evolve. We evaluate DGTN on the public BUSI breast ultrasound dataset using balanced sampling and a joint cross-entropy and Dice loss. The model achieves 68.8% classification accuracy and a Dice score of 0.6227. While its performance remains below that of recent state-of-the-art pipelines, DGTN offers a favorable trade-off between accuracy and computational efficiency within a single unified framework. Ablation experiments indicate that both diffusive attention and gating contribute meaningfully to performance (paired <i>t</i>-test across five cross-validation folds, <i>p</i> < 0.05, with large paired effect sizes, d ≈ 1.0-1.4); because this test is based on only five folds, the result should be interpreted as indicative rather than conclusive, and we report it alongside fold-level effect sizes rather than as a stand-alone confirmation of significance. To the best of our knowledge, this work represents one of the first applications of diffusive graph-transformer co-learning to breast ultrasound imaging, demonstrating the potential of graph-enhanced attention for efficient multi-task medical image analysis.