Generalized Task-Driven Medical Image Quality Enhancement With Gradient Promotion.

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• Task-driven image quality enhancement (IQE) models, like ESTR, leverage mutual enhancement between image processing and visual recognition.
• Existing IQE models often fail to account for the varying and conflicting feature requirements across different vision tasks.

Purpose of the Study:

• To propose a generalized gradient promotion (GradProm) training strategy for task-driven IQE of medical images.
• To address the limitations of current IQE models in handling diverse vision task requirements.

Main Methods:

• A two-sub-model system is proposed: a mainstream image enhancement model and an auxiliary visual recognition model.
• GradProm updates the enhancement model using gradients from both models only when their directions align (cosine similarity).
• If gradients conflict, only the enhancement model's gradient is used, ensuring optimization direction stability.

Main Results:

• Theoretical proof demonstrates that GradProm prevents optimization bias from the auxiliary visual recognition model.
• Extensive experiments on four challenging medical image datasets were conducted.
• GradProm significantly outperformed existing state-of-the-art methods in IQE performance.

Conclusions:

• GradProm offers a robust training strategy for task-driven IQE in medical imaging.
• The method effectively handles varying feature requirements across different vision tasks.
• Superior performance validates GradProm's efficacy over current leading techniques.