Trainable Spatio-Temporal Bilateral Filters: 4D-Filtering for 4DCT Denoising

Schwarz A, Schmidt S, Wohlfahrt P, Dickmann J, Maier A (2025)

Publication Language: English

Publication Type: Conference contribution

Publication year: 2025

Journal

Progress in Biomedical Optics and Imaging - Proceedings of SPIE

Publisher: SPIE

Series: Proceedings of SPIE

Book Volume: 13405

Pages Range: 134054K

Conference Proceedings Title: Medical Imaging 2025: Physics of Medical Imaging

Event location: San Diego, CA US

ISBN: 9781510685888

DOI: 10.1117/12.3047389

Abstract

Time-resolved imaging is crucial for visualizing moving structures such as those affected by cardiac and respiratory motion. In radiation therapy, particularly for thoracic and abdominal tumors, four-dimensional CT (4DCT) is used to track tumors through various breathing states. However, the need to capture data for each intermediate state of the breathing cycle limits the radiation dose per volume, leading to increased image noise. Although the dose is relatively low compared to radiotherapy, repeated 4D scans in fractional radiotherapy can accumulate, potentially discouraging repeat scans. Therefore, denoising 4DCT scans can be beneficial, e specially w hen the temporal dimension of the data can be leveraged. A common challenge for medical imaging methods is the lack of datasets. Most deep learning models require millions of images for training, but this is often unfeasible in the medical domain especially for specialized dataformats like 4DCT. Complex deep learning models can also limit the explainability of the final r esults and in the worst case compromise data integrity through hallucinations. One promising method is 3D trainable bilateral filtering a s i ntroduced b y W agner e t a l., w hich h as s hown s trong p erformance w ith m inimal training data and only few, easily comprehensible parameters. This study aims to expand the trainable bilateral filtering framework into the temporal dimension to better accommodate 4D CT data. Initial evaluations are conducted on paired high-dose and simulated low-dose scans. Both the spatial and combined spatio-temporal filters c onsiderably r educed t he s tandard d eviation o f pixel values in a homogeneous area of the aorta, demonstrating strong denoising performance. A mean SSIM score of 0.95 and PSNR of 41.5 for both indicates successful reconstruction of image content. In images the spatio-temporal filter shows improved conservation of small, low-contrast structures in the lung compared to the spatial filter a lone. T he i mage q uality a ssessment b y t wo m edical i maging e xperts s howed a p reference f or t he 4D filtering i n fi ve ca ses, wh ile th e 3D fil tering was pre ferred in two cas es, in one cas e the y did n't agree.

Authors with CRIS profile

Annette Schwarz Lehrstuhl für Informatik 5 (Mustererkennung) Andreas Maier Lehrstuhl für Informatik 5 (Mustererkennung)

How to cite

APA:

Schwarz, A., Schmidt, S., Wohlfahrt, P., Dickmann, J., & Maier, A. (2025). Trainable Spatio-Temporal Bilateral Filters: 4D-Filtering for 4DCT Denoising. In John M. Sabol, Ke Li, Shiva Abbaszadeh (Eds.), Medical Imaging 2025: Physics of Medical Imaging (pp. 134054K). San Diego, CA, US: SPIE.

MLA:

Schwarz, Annette, et al. "Trainable Spatio-Temporal Bilateral Filters: 4D-Filtering for 4DCT Denoising." Proceedings of the Medical Imaging 2025: Physics of Medical Imaging, San Diego, CA Ed. John M. Sabol, Ke Li, Shiva Abbaszadeh, SPIE, 2025. 134054K.

BibTeX: Download