Perceptual Influence: Improving the Perceptual Loss Design for Low-Dose CT Enhancement

TL;DR

The paper tackles how to design perceptual losses for Low-Dose CT denoising by introducing perceptual influence $\Psi$, a metric that quantifies the perceptual term's contribution to the total loss. By systematically varying the feature level $\phi$ and pretraining context $c$ and calibrating the loss weight $\lambda$ to keep $\Psi \approx 0.95$, the authors demonstrate that conventional LDCT perceptual-loss settings (e.g., $\mathcal{E}$ pretrained on ImageNet with $\phi=block5\_conv4$ and $\lambda\approx0.1$) underperform compared to optimized designs. Results show that ImageNet-pretrained encoders benefit from lower-level features, while domain-specific medical pretraining benefits from higher-level semantic features, leading to improved PSNR, SSIM, and reduced NRMSE without architecture changes. The work provides objective guidelines for configuring perceptual losses in LDCT denoising and releases the source code to support reproducibility and further research.

Abstract

Perceptual losses have emerged as powerful tools for training networks to enhance Low-Dose Computed Tomography (LDCT) images, offering an alternative to traditional pixel-wise losses such as Mean Squared Error, which often lead to over-smoothed reconstructions and loss of clinically relevant details in LDCT images. The perceptual losses operate in a latent feature space defined by a pretrained encoder and aim to preserve semantic content by comparing high-level features rather than raw pixel values. However, the design of perceptual losses involves critical yet underexplored decisions, including the feature representation level, the dataset used to pretrain the encoder, and the relative importance assigned to the perceptual component during optimization. In this work, we introduce the concept of perceptual influence (a metric that quantifies the relative contribution of the perceptual loss term to the total loss) and propose a principled framework to assess the impact of the loss design choices on the model training performance. Through systematic experimentation, we show that the widely used configurations in the literature to set up a perceptual loss underperform compared to better-designed alternatives. Our findings show that better perceptual loss designs lead to significant improvements in noise reduction and structural fidelity of reconstructed CT images, without requiring any changes to the network architecture. We also provide objective guidelines, supported by statistical analysis, to inform the effective use of perceptual losses in LDCT denoising. Our source code is available at https://github.com/vngabriel/perceptual-influence.

Figures (5)

• Figure 2: Perceptual influence $\Psi(\lambda, \phi, c)$ as a function of the loss weight $\lambda$, computed for each experimental configuration $\{E_1, E_2, E_3, E_4\}$. The curves illustrate how the contribution of the perceptual term $\ell_\text{PL}$ evolves relative to the total cost $J_\theta$ across $\lambda$ values from $10^{-7}$ to $1$, under different combinations of feature representations $\phi$ and pre-training contexts $c$.
• Figure 3: Error heatmaps between denoised outputs and ground truth for each experimental configuration.
• Figure : (a)
• Figure : (a)
• Figure : (b)