A Comparative Analysis of Loss Functions in Deep Learning-Based Inverse Problems

Type: Project

Status: running

Supervisors: Yipeng Sun, Linda-Sophie Schneider, Andreas Maier


In recent years, deep learning has emerged as a transformative force in the realm of image processing, particularly in addressing inverse problems such as denoising and artifact reduction in medical imaging. This research aims to systematically investigate the impact of various loss functions on deep learning-based solutions for inverse problems, with a focus on low-dose Computed Tomography (CT) imaging.

Low-dose CT, while beneficial in reducing radiation exposure, often suffers from increased noise and artifacts, adversely affecting image quality and diagnostic reliability. Traditional denoising techniques, although effective to some extent, struggle to maintain a balance between noise reduction and the preservation of crucial image details. Deep learning, especially Convolutional Neural Networks (CNNs), has shown promising results in surpassing these traditional methods, offering enhanced image reconstruction with remarkable fidelity.

However, the choice of loss function in training deep learning models is critical and often dictates the quality of the reconstructed images. Commonly used loss functions like Mean Squared Error (MSE) or Structural Similarity Index (SSIM) have their limitations and may not always align well with human perceptual quality. This research proposes to explore and compare a variety of loss functions, including novel and hybrid formulations, to evaluate their efficacy in enhancing image quality, reducing noise, and eliminating artifacts in low-dose CT images.