ABSTRACT NUMBER: ESOC2026A533 DEEP LEARNING BASED STREAK ARTEFACT REDUCTION IN CEREBRAL CONE-BEAM CT

Abstract Background and aims Streak artefacts in cerebral cone-beam CT (CBCT) can compromise diagnostic assessment and treatment planning by obscuring fine neuroanatomy. We evaluated two deep learning-based (DL) post-processing models for streak artefact suppression and diagnostic image quality enhancement, and compared them with original unprocessed images. Methods To evaluate the DL-based models (E-01 and E-02), trained to suppress streak artefacts in reconstructed CBCT images, we conducted a reader study (n=9, average 10.5 years of cerebral CBCT imaging experience). Readers assessed 26 unseen clinical cases with varying artefact levels (none to severe), rating streak severity, sharpness, and overall image quality on 5-point Likert scales for unprocessed versus DL-processed images. Quantitative fidelity on held-out paired test data was measured using MAE, SSIM, and PSNR. Results Both models reduced artefact severity and improved perceived image quality versus unprocessed scans. Unfavourable streak-severity ratings fell from 55% (unprocessed) to 36% (E-01) and 26% (E-02). For scans with at least mild baseline streaking, unfavourable overall-quality ratings decreased from 25% to 14% (E-01) and 8% (E-02). E-02 provided the strongest suppression but caused modest blurring (unfavourable sharpness 15% vs 7% for unprocessed and E-01). Despite E-02's lower fidelity (MAE 0.536 vs 0.366; SSIM 0.9894 vs 0.9957; PSNR 39.32 vs 42.46), readers preferred it for diagnostic utility. Conclusions We demonstrate DL-based post-processing potential to enhance CBCT interpretability by reducing streak artefacts. Readers preferred more aggressive suppression despite modest blurring, indicating that standard image-fidelity metrics may not reliably capture clinically relevant usefulness in artefact-reduction tasks. Conflict of interest Martin Kremnický: nothing to disclose, Luis Albert Zavala-Mondragón: nothing to disclose, Fred van Nijatten: nothing to disclose, Erik Hummel: nothing to disclose, Danny Ruijters: nothing to disclose, Nikolas Schnellbächer: nothing to disclose, Fons van der Sommen: nothing to disclose, Nicole Cancelliere: nothing to disclose, Vitor Mendes Pereira: nothing to disclose Figure 1 - belongs to Background and aims Figure 2 - belongs to Methods Figure 3 - belongs to Results Figure 4 - belongs to Conclusions