Introduction
Computed Tomography (CT) is a cornerstone of modern diagnostic medicine, yet the ionizing radiation involved in standard scans poses potential health risks. Low-dose CT (LDCT) protocols are safer but introduce significant noise and artifacts that can obscure clinical details. Traditional deep learning denoising methods often rely on pixel-level losses like Mean Squared Error (MSE), which frequently result in "over-smoothed" or "waxy" images where fine textures are lost. Furthermore, these models typically treat all pixels identically, ignoring the distinct noise characteristics and semantic structures of different human tissues. This paper introduces ASCON, an anatomy-aware framework designed to preserve structural integrity and provide interpretable feature representations.
Methodology
The ASCON framework addresses the limitations of standard denoising through two primary innovations: an efficient attention-based architecture and a multi-scale contrastive learning strategy.
ESAU-Net: Efficient Self-Attention-based U-Net
To capture global context without the prohibitive computational costs of standard spatial self-attention, ASCON utilizes ESAU-Net. This architecture employs channel-wise self-attention, computing cross-covariance across feature channels rather than spatial locations. This allows the model to maintain a global receptive field while remaining memory-efficient for high-resolution medical images.
MAC-Net: Multi-scale Anatomical Contrastive Network
The "anatomy-aware" capability is driven by MAC-Net, which operates at two distinct scales to align the denoised output with ground-truth normal-dose CT (NDCT) images:
- Patch-wise non-contrastive module: Inspired by the BYOL (Bootstrap Your Own Latent) strategy, this module aligns high-level anatomical semantics between denoised patches and NDCT patches.
- Pixel-wise contrastive module: Utilizing a local InfoNCE loss, this module ensures pixel-level consistency while distinguishing between different semantic structures, effectively pushing the model to recognize tissue boundaries.
The framework is optimized using an alternate learning strategy, where the denoiser (ESAU-Net) and the contrastive network (MAC-Net) are updated iteratively, similar to the training process of Generative Adversarial Networks (GANs).
Results and Interpretability
ASCON was evaluated on the Mayo-2016 and Mayo-2020 benchmarks, outperforming state-of-the-art models such as RED-CNN and DU-GAN. Quantitatively, ASCON achieved a PSNR of 44.48 dB on the Mayo-2016 dataset, a notable improvement over standard U-Net baselines. Qualitatively, the framework demonstrates superior edge preservation and noise suppression, particularly in low-contrast lesions where it maintains sharper boundaries between blood vessels and surrounding organs.
Beyond performance metrics, ASCON provides a unique level of anatomical interpretability. Without explicit segmentation labels, the contrastive loss forces the model to learn distinct feature representations for different tissues. As shown in the t-SNE visualizations, ASCON's internal features naturally cluster into recognizable anatomical structures, whereas standard models produce much more disorganized feature spaces.
Conclusion
ASCON represents a significant step toward biology-informed deep learning in medical imaging. By incorporating anatomical priors through multi-scale contrastive learning, the framework achieves high-fidelity denoising while offering transparency into how the model perceives different tissues. Future work includes extending the framework to 3D volumetric data and conducting further studies to validate its clinical utility for radiologists.
Summary updated