Deep Learning Ring Artifact Removal in Computed Tomography: A Two-Head U-Net Benchmarked Against Classical Radial profile Filtering

Ring artifacts, caused by detector pixels whose gain or offset is slightly wrong, appear in reconstructed computed tomography (CT) slices as concentric circles around the rotation centre. They can hide small features and break threshold-based segmentation. Classical ring removal relies on hand-tuned filters, and real scans have no ground truth for measuring how well a method works. This project builds a synthetic CT pipeline (phantom, parallel-beam projection, Poisson noise, detector gain and offset error, filtered back-projection) in which the true ring is known for every image, and trains a two-head U-Net that predicts the ring image, which is subtracted from the input, and a ring mask. On 1000 training, 100 validation and 100 test slices of 256 × 256 pixels, where the test slices come from a phantom family never seen in training, the U-Net raises the peak signal-to-noise ratio (PSNR) from 33.11 dB to 39.54 dB and the structural similarity (SSIM) from 0.868 to 0.957. A radial-profile classical baseline tuned on the validation set gains 0.77 dB on validation but loses 0.56 dB on the unseen family. The U-Net gain depends on ring type, from +13.9 dB for dead_hot rings to +0.4 dB for faint band rings. Training with 100, 500 and 1000 images gives 35.29, 38.42 and 39.54 dB; three random seeds give 39.40 ± 0.15 dB; removing the Dice term from the loss raises PSNR to 39.89 dB but lowers mask Dice from 0.640 to 0.605. The mask head is the weaker output (Dice 0.640 overall, 0.07 on band rings). Because all data are simulated, the sim-to-real gap remains the main open question.

Authors

Institutions

Publication Details

Journal
Research Square
Published
2026-11-18
DOI
https://doi.org/10.21203/rs.3.rs-11739530/v1
Primary Topic
Advanced X-ray and CT Imaging
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

Deep Learning Ring Artifact Removal in Computed Tomography: A Two-Head U-Net Benchmarked Against Classical Radial profile Filtering

Ekta Chahal, Mayank Goswami
Research Square
Advanced X-ray and CT Imaging
preprint

Deep Learning Ring Artifact Removal in Computed Tomography: A Two-Head U-Net Benchmarked Against Classical Radial profile Filtering

Ekta Chahal, Mayank Goswami
preprint en

Abstract

Ring artifacts, caused by detector pixels whose gain or offset is slightly wrong, appear in reconstructed computed tomography (CT) slices as concentric circles around the rotation centre. They can hide small features and break threshold-based segmentation. Classical ring removal relies on hand-tuned filters, and real scans have no ground truth for measuring how well a method works. This project builds a synthetic CT pipeline (phantom, parallel-beam projection, Poisson noise, detector gain and offset error, filtered back-projection) in which the true ring is known for every image, and trains a two-head U-Net that predicts the ring image, which is subtracted from the input, and a ring mask. On 1000 training, 100 validation and 100 test slices of 256 × 256 pixels, where the test slices come from a phantom family never seen in training, the U-Net raises the peak signal-to-noise ratio (PSNR) from 33.11 dB to 39.54 dB and the structural similarity (SSIM) from 0.868 to 0.957. A radial-profile classical baseline tuned on the validation set gains 0.77 dB on validation but loses 0.56 dB on the unseen family. The U-Net gain depends on ring type, from +13.9 dB for dead_hot rings to +0.4 dB for faint band rings. Training with 100, 500 and 1000 images gives 35.29, 38.42 and 39.54 dB; three random seeds give 39.40 ± 0.15 dB; removing the Dice term from the loss raises PSNR to 39.89 dB but lowers mask Dice from 0.640 to 0.605. The mask head is the weaker output (Dice 0.640 overall, 0.07 on band rings). Because all data are simulated, the sim-to-real gap remains the main open question.

Research Square
Indian Institute of Technology Roorkee (IN)
Advanced X-ray and CT Imaging
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon