Dual Polarization Modality Fusion Network for Assisting Pathological Diagnosis
IEEE Transactions on Medical Imaging
|September 26, 2022
Summary
This study introduces a dual polarization modality fusion network (DPMFNet) to enhance cancer diagnosis by combining polarization imaging features. The DPMFNet improves pathological diagnosis, especially for small datasets and low-resolution images.
Area of Science:
- Medical imaging
- Computational pathology
- Biomedical optics
Background:
- Polarization imaging offers rich optical and microstructural information from cancer tissues.
- Integrating polarization data to improve pathological diagnosis remains a challenge.
Purpose of the Study:
- To develop a novel network, the dual polarization modality fusion network (DPMFNet), for enhanced pathological diagnosis.
- To effectively fuse and utilize information from dual polarization imaging modalities.
Main Methods:
- A multi-stream convolutional neural network (CNN) structure with a switched attention fusion module.
- Dual-polarization contrastive training to synthesize and align feature representations.
- Utilizing Grad-CAM for visualization and analysis of important image regions.
Main Results:
- The DPMFNet demonstrated superior performance in assisting pathological diagnosis across three cancer datasets.
- Significant improvements were observed particularly in small datasets and low imaging resolution scenarios.
- Grad-CAM analysis confirmed the complementary roles of polarization and pathological images in diagnosis.
Conclusions:
- The DPMFNet effectively integrates dual polarization imaging features for enhanced cancer diagnosis.
- This approach shows potential for improving pathological aided diagnosis and advancing digital pathology.
- The switched attention mechanism and contrastive training contribute to improved feature representation and semantic relatedness.
