Anatomically adaptive feature-wise linear modulation for deep learning-based low-dose CT denoising

Abstract

Objective. Low-dose computed tomography (LDCT) reduces radiation dose but, introduces heterogeneous noise due to different photon attenuation based on anatomical tissue. Most deep learning techniques assume uniform noise in LDCT and perform equal noise removal across different regions, leading to sub-optimal performance across different tissues. This work aims to design a physics-based framework that explicitly models region-dependent noise characteristics to improve LDCT noise removal. Approach. We propose an anatomically adaptive noise reduction framework. The proposed anatomically adaptive feature-wise linear modulation (FiLM) model consists of a U-Net architecture that integrates two complementary units: the global FiLM unit in the encoder, which modifies features globally based on global image statistics to remove overall noise, and the local FiLM unit in the decoder, which modifies features locally based on the type of anatomical tissue. This dual design enables the modeling of overall image noise characteristics in addition to the removal of local noise associated with each anatomical tissue. Main results. The model performance was evaluated using AAPM–Mayo Clinic LDCT dataset, and the trained model tested on the TCIA dataset. The proposed model outperformed all competing methods. Local noise analysis showed that noise removal was consistent across different anatomical regions, achieving 37.14% in the lung, 48.75% in soft tissue, and 35.18% in bone. Visual results also confirmed significant improvements in noise removal, preservation of structural details, and reduction of non-residual distortions. Furthermore, the model demonstrated its ability to generalize under domain shift. Significance. This work presents a framework for anatomically adapted noise removal by linking feature modification with the physical properties of noise in LDCT through a dual-modulation process for both general and tissue-related noise. The model achieves a balance between noise removal and preservation of anatomical detail, making it a robust approach to LDCT noise removal.