GAN-Guided Few-Shot Attention Network for Medical Images Fusion Quality Assessment

Area of Science:

• Medical imaging
• Artificial intelligence in healthcare
• Image processing

Background:

• Medical image fusion (MIF) is crucial for diagnostics and treatment planning.
• Medical image fusion quality assessment (MIFQA) is vital for enhancing MIF performance.
• Acquiring medical reference images for MIFQA is challenging due to data scarcity and the need for expert knowledge.

Purpose of the Study:

• To develop a novel two-stage model for MIFQA that overcomes the limitations of traditional methods.
• To reduce the dependence on extensive medical prior knowledge and reference images in MIFQA.
• To improve the accuracy and efficiency of quality assessment for medical image fusion.

Main Methods:

• A Generative Adversarial Network (GAN)-based Quality-aware Network (QANet) was designed to generate reference images based on radiologist scores.
• A class attention siamese network (CASNet) utilizing class activation mapping (CAM) was employed for few-shot learning.
• The model focuses on key lesion areas to effectively assess image quality with limited data.

Main Results:

• The proposed two-stage model successfully generates quality-specific reference images.
• CASNet effectively utilizes limited reference images by focusing on critical regions.
• Experimental results on a custom MIFQA dataset demonstrate superior performance compared to state-of-the-art methods.

Conclusions:

• The developed two-stage model offers a promising solution for MIFQA, addressing the challenge of limited reference data.
• The method enhances the focus on diagnostically relevant areas, leading to more accurate quality assessments.
• This approach has the potential to significantly advance precision diagnostics through improved medical image fusion quality assessment.