Work

Bottleneck Width, Loss Design and the Limits of Skip-Free Convolutional Denoising Autoencoders for Simulated Low-Dose CT

Year: 2026
Type: preprint
Abstract: Convolutional denoising autoencoders (DAEs) are attractive for computed tomography (CT) because they offer a small model and a compact, reusable latent code, but it is unclear whether a skip-free DAE can compete with feed-forward residual networks. We examine this question on 41 abdominal CT volumes from the Medical Segmentation Decathlon (820 real slices, split by volume) and 300 synthetic multi-material phantoms, corrupted with a two-stage noise model (Poisson photon noise followed by Gaussian read noise). We train convolutional DAEs with bottleneck widths of 8 to 256 channels, a variational variant and a composite MSE+SSIM loss, and compare them with a DnCNN-style network and with Gaussian, median and non-local-means filters on identical held-out volumes. DnCNN reached 44.21 dB PSNR and 0.979 SSIM, against 35.07 dB and 0.938 for the best MSE-trained autoencoder; the noisy input scored 35.56 dB and 0.756. The composite loss raised the autoencoder to 35.65 dB and 0.950. Quality was flat between 32 and 256 channels (PSNR spread 0.22 dB) and dropped only at 16 and 8 channels (− 0.9 and − 2.5 dB relative to width 128). On real slices the width-128 autoencoder lowered PSNR relative to the noisy input in 71% of cases, mainly at thin, high-contrast structures, even though SSIM improved by about 0.12. Latent codes predicted phantom material with 66–84% accuracy (chance 20%) but formed no clusters (silhouette ≈ 0). The autoencoder needed about 34 times less compute per tile than DnCNN, yet was far less accurate. In this setting the reconstruction-through-a-bottleneck design, rather than bottleneck size, appears to limit denoising quality; the role of skip connections remains to be isolated.
Language: English
Open Access status: green
OpenAlex ID: W7233925065
Cites: 26
Cited by:
Related to: 10