An ultrasound spectrum-to-graph (US2G) framework with multi-feature graph attention networks for automated lesion classification in B-mode ultrasound imaging
Year: 2026
Type: article
Abstract: Automated classification of lesions in B-mode ultrasound images remains challenging due to speckle noise, irregular lesion boundaries, and variable echo intensity distributions. Existing convolutional neural networks (CNNs), Vision Transformers (ViTs), and ultrasound foundation models share a critical limitation: all represent regions of interest (ROIs) as regular pixel grids, failing to exploit the irregular spatial topology of acoustic energy diffusion patterns that encode diagnostically relevant tissue microstructure. We propose US2G , a three-stage spectrum-to-graph framework that (i) extracts an adaptive Acoustic Energy Diffusion Region (AEDR) via multi-level Otsu thresholding, without manual annotation; (ii) constructs a multi-feature weighted graph with six heterogeneous node features and a composite three-term edge similarity function combining intensity affinity, spatial proximity, and textural coherence; and (iii) classifies variable-topology graphs using edge-conditioned multi-head GATv2 with differentiable hierarchical pooling. Experiments on the BUSI (breast), DDTI (thyroid), and CAMUS (cardiac) datasets yield classification accuracies of 91.47%, 93.82%, and 89.63%, respectively, corresponding to absolute gains of 5.60, 8.82, and 6.13 percentage points over the strongest CNN baseline and 3.97, 5.49, and 3.88 points over a GCN operating on an identical graph. The proposed US2G achieves this with 48.7 K parameters and 6 ms per-ROI inference, more than two orders of magnitude below the parameter count of contemporary ultrasound foundation models. Attention visualisations align with radiologist-interpretable sonographic features, supporting clinical interpretability. We also explicitly delineate the assumptions and applicability constraints of the framework, including its dependence on ROI quality, its single-frame scope, and the limited scale of the benchmarks currently available.
Authors: Arshad Iqbal, Sohail Asghar
Institutions: Pakistan Academy of Sciences, Air University, HITEC University
Language: English
Topic: AI in cancer detection
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
Estimated APC paid: $2,590
OpenAlex ID: W7222798671