Liver fibrosis MR images classification based on higher-order interaction and sample distribution rebalancing

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Biomedical Engineering

Background:

• Liver fibrosis MR images present challenges due to fragmented fractal features and poor interconnectivity, leading to incomplete learning and low accuracy.
• Image quality is further degraded by light scattering, quantum noise, and artifacts like breathing during MRI acquisition.

Purpose of the Study:

• To improve the recognition accuracy of liver fibrosis in MR images.
• To enhance feature learning and capture sample characteristics more effectively using advanced deep learning techniques.

Main Methods:

• Implemented recursive gated convolution within the ResNet18 network to capture higher-order spatial information interactions.
• Introduced the Adaptive Rebalance loss function to expand inter-class distance and narrow intra-class differences.
• Incorporated a feature paradigm as a learnable adaptive attribute into the angular margin auxiliary function.

Main Results:

• The proposed method demonstrated an average improvement of two percent in recognition accuracy compared to the standard ResNet18.
• Enhanced correlation between features and improved focus on pixel-level dependencies for global image interpretation.
• The Adaptive Rebalance loss function effectively improved the model's discriminative ability.

Conclusions:

• The integration of recursive gated convolution and Adaptive Rebalance loss function significantly enhances liver fibrosis MR image classification.
• The developed model offers a more robust and accurate approach to analyzing liver fibrosis from MR images, addressing limitations of existing methods.