S2039
Physics - Image acquisition and processing
ESTRO 2026
Digital Poster Highlight 62
Compression of Cine-MRI data and its impact on AI- based target tracking in MR-guided radiotherapy Tom Julius Blöcker Department of Radiation Oncology, LMU University Hospital, Munich, Germany Purpose/Objective: Magnetic resonance image-guided radiotherapy (MRgRT) typically relies on 2D+t CineMRI to monitor motion. With increasing adoption and higher imaging frequencies, CineMRI produces growing datasets, creating challenges for storage, transfer, and large- scale research use. Current workflows mainly use standard lossless compression (zlib/DEFLATE), which provides only limited reduction in file sizes. This study evaluates lossless and lossy compression methods for CineMRI, and evaluates how lossy compression artefacts influence the accuracy of AI-based target tracking. Material/Methods: The publicly available TrackRAD2025 (1) dataset (sagittal, 16-bit CineMRI frames) was compressed using a newly developed tool and format extending the MHA format to support compression methods beyond zlib.Three compression scenarios were evaluated: (a) lossless at 16-bit, (b) lossless after quantisation to 8-bit, and (c) lossy compression at 8- bit. Algorithms included entropy-based (zlib, bzip2, Brotli, Zstandard) and image-based methods, with intraframe-coding (H.264/AVC, H.265/HEVC) and without (JPEG, WebP, JPEG XL).Compression ratios (CR = uncompressed/compressed size) were calculated. For lossy methods, fidelity was assessed with peak signal-to-noise ratio (PSNR).To assess downstream utility, decompressed CineMRIs were used with AI- based target tracking models. Segmentation accuracy was evaluated with the Dice similarity coefficient (DSC) and Euclidean center distance (ECD), and robustness
Model performance depended on PSNR, but not on the specific compression method (JPEG XL vs. H.265). At PSNR above 45 dB, segmentation accuracy was preserved ( Δ DSC <0.01, Δ ECD <0.05 mm), with mean compression ratios of 22 (JPEG XL) and 60 (H.265). Noticeable performance loss emerged only below 40 dB, with possible compression ratios of 41 and 172, respectively.
was quantified as the absolute change in DSC compared to uncompressed data (| Δ DSC|). Results:
For lossless 16-bit compression, JPEG XL provided the highest compression ratio (mean CR 4.4), substantially improving on zlib (CR 2.5). 8-bit quantisation halved dataset size with little degradation (PSNR ≈ 55 dB), boosted the efficiency of all compression methods and enabled use of video codecs. Again, JPEG XL produced the highest compression ratios (CR 4.8).For lossy 8-bit compression, video-based methods delivered the most favorable balance between compression and fidelity (CR-to-PSNR).
Conclusion: Advanced methods can improve CineMRI
compression, either losslessly for archival purposes (CR 4.4-4.8 using JPEG XL vs 2.5 using zlib/DEFLATE) or lossy (using H.265 at much higher CRs) when moderate artefacts are acceptable.AI-based target tracking models remain robust above ~45 dB PSNR, establishing a practical threshold for efficient large- scale CineMRI handling research workflows.
Made with FlippingBook - Share PDF online