EAGLE: An Edge-Aware Gradient Localization Enhanced Loss for CT Image Reconstruction

TL;DR

The paper tackles the loss-function design challenge in CT image reconstruction, where traditional MSE-based objectives blur fine details. It introduces Eagle-Loss, which leverages localized gradient variance and frequency-domain analysis to preserve edges and high-frequency information, and demonstrates its effectiveness as both a main loss and a regularizer. Through experiments on low-dose CT reconstruction and field-of-view extension across multiple networks, Eagle-Loss yields improved visual sharpness and edge fidelity compared to state-of-the-art losses, with public code provided for replication. The approach is model-agnostic and has potential for broader impact in medical imaging and beyond, contingent on careful hyperparameter tuning.

Abstract

Computed Tomography (CT) image reconstruction is crucial for accurate diagnosis and deep learning approaches have demonstrated significant potential in improving reconstruction quality. However, the choice of loss function profoundly affects the reconstructed images. Traditional mean squared error loss often produces blurry images lacking fine details, while alternatives designed to improve may introduce structural artifacts or other undesirable effects. To address these limitations, we propose Eagle-Loss, a novel loss function designed to enhance the visual quality of CT image reconstructions. Eagle-Loss applies spectral analysis of localized features within gradient changes to enhance sharpness and well-defined edges. We evaluated Eagle-Loss on two public datasets across low-dose CT reconstruction and CT field-of-view extension tasks. Our results show that Eagle-Loss consistently improves the visual quality of reconstructed images, surpassing state-of-the-art methods across various network architectures. Code and data are available at \url{https://github.com/sypsyp97/Eagle_Loss}.

Figures (5)

• Figure 1: Illustration of steps involved in computing the magnitude spectrum $M_x(I)$ and $M_y(I)$ from an image $I$. In this figure, larger patches are used for better visualization.
• Figure 2: (a) The structure of TF-FBP. Here $H$ represents the trainable filter, $A^{-1}$ represents the differentiable backprojection operator, while $F^{-1}$ and $F$ denote the inverse and forward . The differentiable backprojection operator was implemented using PYRO-NN syben2019pyro. (b) RED-CNN with its encoder-decoder structure. The input of the encoder is the reconstructed image using .
• Figure 3: Comparison of our method with , sun2024data, gatys2016image, abrahamyan2022gradient, and Perceptual Loss johnson2016perceptual. (a) Visualization results for low-dose CT reconstruction on LoDoPaB-CT dataset. The Red and blue dashed boxes present RED-CNN and TF-FBP. Input image of RED-CNN is the reconstruction from algorithm with a Ramp filter. (b) Visualization of extension results on SMIR dataset. The blue dotted circle denotes the original boundary.
• Figure 4: Comparison of Eagle-Loss regularization versus regularization and no regularization in the reconstruction on LoDoPaB-CT dataset.
• Figure 5: Visual representation of the filters and their corresponding reconstructed images at varying $\kappa$ values in the TF-FBP model. This illustration highlights the direct influence of the cutoff frequency on the enhancement of high-frequency details in image reconstruction.