CT‐Specific FFT‐Augmented LaMa Inpainting: Adaptation, Validation, and Clinical Considerations

ABSTRACT

This study presents a CT‐specific adaptation of the FFT‐enhanced large mask inpainting (LaMa) framework for restoring corrupted thoracic CT images. Instead of introducing a new architecture, the proposed approach tailors the fast Fourier convolution (FFC)‐based LaMa model through anatomy‐guided edge masking, CT‐specific intensity normalisation and an increased number of residual blocks. A dataset of 2100 thoracic CT slices (256 × 256 pixels) from 70 patients was used for model development and validation. Images were preprocessed using bilinear interpolation and min–max normalisation. A Canny edge–based masking strategy with approximately 20% mask coverage simulated selective spatial information reduction to evaluate reconstruction robustness. The encoder–decoder network employed eight FFC residual blocks and a hybrid loss function combining L ₁ and VGG‐16 perceptual losses. Training used the Adam optimiser with patient‐level five‐fold cross‐validation, a batch size of 16, a learning rate of 1 × 10 ^{− 3} and 50 epochs. The proposed model achieved a mean PSNR of 34.07 ± 1.43 dB and SSIM of 0.9437 ± 0.0367, significantly outperforming both Mamba and greyscale‐adapted LaMa baselines while reducing training time by approximately 25%. A preliminary radiologist evaluation yielded favourable visual quality scores. These results demonstrate the effectiveness of CT‐specific adaptation for medical image inpainting and suggest potential applications in storage‐aware CT imaging, warranting further validation against established compression methods.