ESTRO 2025 - Abstract Book

S2965

Physics - Image acquisition and processing

ESTRO 2025

Material/Methods: We evaluated CPT-DIR on 37 head-and-neck patient images, of which 7 patients with proton treatment plans were used for dose warping evaluation. For each patient, bone contours were extracted by TotalSegmentator [4] , and other OAR contours were delineated by physicians. One patient with its CT image was randomly selected as the fixed image to define the common coordinate system (CCS). CPT-DIR was then applied between the fixed CCS image and all other patient images respectively. Region-specific constraints, such as dice similarity of brainstem, were incorporated into the CPT-DIR optimization, enabling the framework to use both image features and organ/region correspondence, thereby improving registration accuracy in the low-contrast, dosimetrically relevant regions. The Dice Similarity Coefficient (DSC) and Dose-Organ Overlap (DOO) scores were used to evaluate the geometric accuracy of the inter-patient registration. To evaluate the effectiveness of contour-informed constraints, we compared registration results with and without brainstem contours as input, while all DIR registrations used the bone contour as the default input. CPT-DIR 's performance was systematically compared to the conventional B splines DIR with all above matrices.

Results: Across DIR results for 36 patients, CPT-DIR achieved a mean absolute error (within body) of 100.06 ± 12.01 HU compared to CCS, and 155.90 ± 15.10 HU for B-splines. Figure 2 summarizes subset corresponding image differences and structure DSC/DOO, demonstrating that CPT-DIR outperforms the baseline B-splines DIR across all