TDSF-Net: Tensor Decomposition-Based Subspace Fusion Network for Multimodal Medical Image Classification

Area of Science:

• Medical Imaging
• Deep Learning
• Computer Vision

Background:

• Multimodal data offers complementary information for medical image classification.
• Existing fusion methods struggle with high dimensionality and computational complexity.
• Lack of effective fusion strategies hinders deep learning model performance.

Purpose of the Study:

• To propose a novel subspace fusion network utilizing tensor decomposition (TD) for enhanced multimodal medical image classification.
• To address the limitations of conventional data fusion techniques in handling multimodal medical data.
• To improve the accuracy and efficiency of deep learning models in classifying medical images from multiple sources.

Main Methods:

• Developed a Tucker low-rank tensor decomposition (TD) module to map high-dimensional features into a low-rank subspace, reducing redundancy.
• Implemented a cross-tensor attention mechanism to fuse features within the subspace, enhancing representational ability and inter-component interactions.
• Evaluated the proposed Tensor Decomposition Subspace Fusion Network (TDSFNet) on diverse multimodal medical image datasets.

Main Results:

• The proposed TDSFNet significantly outperforms state-of-the-art (SOTA) methods in multimodal medical image classification tasks.
• Demonstrated effectiveness and generalization ability across one self-established and three public datasets.
• Achieved superior performance by efficiently handling multimodal data and high-dimensional features.

Conclusions:

• The subspace fusion network with tensor decomposition offers a powerful approach for multimodal medical image classification.
• The method effectively reduces computational complexity and improves feature representation.
• The proposed TDSFNet provides a robust and generalizable solution for enhancing deep learning in medical imaging analysis.