BgNet

BgNet classifies benign and malignant spinal tumors by integrating axial and sagittal MRI planes with patient-level clinical information using a multi-plane attention learning framework.

Key Features:

• Multi-Plane Attention Learning: Leverages attention across axial and sagittal MRI planes to match and correlate tumor-area features between planes for robust feature extraction.
• Bipartite Graph Structure: Represents each tumor area in a plane as a vertex and encodes matches between areas across planes as edges to enable graph-based multi-plane integration.
• Feature Fusion Model (ResNetST): Combines ResNet convolutional feature extractors with Swin-Transformer visual attention modules to capture both local and global correlation features.
• Modular Architecture: Implements five modules—multi-plane fusion (bipartite graph), input-layer fusion, feature-layer fusion, decision-layer fusion, and output module—for multi-level data fusion.
• Patient-level Clinical Integration: Incorporates patient-level clinical information alongside multi-plane MRI data for patient-level diagnosis.

Scientific Applications:

• Spinal tumor classification: Distinguishes benign versus malignant spinal tumors using axial and sagittal MRI combined with clinical data.
• Performance evaluation and clinician augmentation: Validated on 430 cases (297 training, 133 testing) with overall accuracy (ACC) 79.7%, matching or exceeding four doctors (D1: 70.7%, D2: 54.1%, D3: 79.7%, D4: 72.9%) and improving doctors' ACC by 4.5% (D1), 21.8% (D2), 0.8% (D3), and 3.8% (D4).

Methodology:

Tumor areas are spliced from different plane images at the input layer; features are extracted using the ResNetST model (ResNet + Swin-Transformer); multi-plane attention learning is implemented via a bipartite graph and fused across input, feature, and decision layers using the five-module architecture.