Small-lesion image encoder
Extends PENet with 3D channel attention, spatial attention, and depth-wise convolution fusion to capture subtle lesion regions across volumetric slices.
Paper 2709 / Medical AI
Small Lesions-aware Bidirectional Multimodal Multiscale Fusion Network for Lung Disease Classification
MMCAF-Net targets a practical clinical diagnosis problem: small lesions can disappear inside 3D medical images, while imaging and electronic health record features are difficult to align directly.
Abstract distilled
Multimodal fusion built for small lesions and uneven feature scales.
The paper proposes a framework that combines 3D medical image features and clinical table data. It introduces E3D-MSCA for lesion-specific multiscale image encoding, MSCA for cross-modal integration, and BSF for reducing uncertainty when fusing features across scales.
Clinical table encoder
Encodes tabular attributes such as gender, age, weight, TNM stage, and smoking history before multimodal alignment.
Cross-modal fusion
Uses multiscale cross attention and bidirectional scale fusion to reduce dimensional conflict between image and clinical features.
Lung-PET-CT-Dx result
Higher classification reliability on most reported metrics.
On the public Lung-PET-CT-Dx dataset, MMCAF-Net outperformed comparative multimodal approaches on ACC, F1, specificity, sensitivity, PPV, and NPV. The study reports a lower false-positive tendency through a stronger PPV score.
ACC0.791
F10.690
Specificity0.857
Sensitivity0.667
PPV0.714
NPV0.828
Selected comparison
MethodACCF1PPV
PECon0.7440.6450.625
MedFuse0.7210.5710.615
MMTM0.6980.5810.562
DAFT0.7290.6670.684
MMCAF-Net0.7910.6900.714