289 - Impact of Photon-Counting CT on Organ Segmentation Quality

Purpose/Objective(s):

Robust and accurate organ segmentation is crucial for precise automated radiotherapy (RT) planning. With the advent of photon-counting CT (PCCT) in RT, a broad range of image contrasts beyond conventional 120 kVp single-energy CT (SE) and quantitative material information are always available. This study investigated the impact of different PCCT image impressions on the stability of a commercially available automatic organ segmentation algorithm.

Materials/Methods:

PCCT scans of 274 oncological patients were acquired in native, arterial, and/or venous iodine contrast phases and reconstructed in a resolution of 0.49×0.49×1 mm³, various sharpness levels (Qr32/smooth–Qr76/sharp) and image contrasts covering virtual monoenergetic images (40–190 keV), iodine (IOD), virtual non-contrast (VNC), stopping-power ratio (SPR), electron density (ED), and effective atomic number (EAN). Iterative metal artifact reduction (iMAR) was applied for cases with metal implants. The conformity of 182 organs automatically segmented on each dataset was assessed by Dice coefficient (DC) and surface Dice (SDC, 1.5 mm) with respect to a SE-like image (70 keV VMI, Qr40).

Results:

Organ conformity varied mainly with image contrast. Low-energy VMIs (high contrast, more pronounced metal artifacts) resulted in larger variability, whereas higher energies (lower contrast, reduced metal artifacts) showed higher robustness (Table 1). SPR and ED performed comparably to 60 keV VMI, whereas IOD and EAN reached a median DC below 70%. High-contrast structures (lung, bone) were less affected than organs in low-contrast regions (lymph nodes, soft tissue, nerves). The impact of low-energy VMIs on conformity approached the magnitude of metal artifacts. The omission of iMAR for hip implants led to a median DC reduction of 15% for prostate. Varying image sharpness had only minor influence with median DC >98%.

Conclusion:

Various image contrasts provided by PCCT substantially influenced segmentation conformity. To ensure high-quality organ segmentations, image reconstructions should be representative for image contrasts used in network training to overcome its sensitivity to unfamiliar contrast domains. This finding indicates that quantitative spectral information inherent to PCCT may represent an underexplored resource to enhance segmentation accuracy and robustness.

Abstract 289 - Table 1: Median DC±SD, SDC in Parentheses.

Table 1: Median DC±SD, SDC in parentheses.

Organ	40 keV	60 keV	190 keV	VNC
Overall	92±20 (91) %	98±8 (100) %	94±12 (95) %	93±13 (94) %
Lung	100±5 (99) %	100±4 (100) %	100±4 (100) %	100±7 (100) %
Pelvic Bone	97±2 (98) %	99±1 (100) %	97±2 (100) %	97±5 (99) %
Larynx	90±17 (91) %	97±3 (100) %	93±6 (94) %	92±14 (94) %
Brainstem	89±17 (80) %	97±4 (100) %	94±5 (96) %	93±12 (94) %
Parotid	86±17 (71) %	97±12 (99) %	93±13 (88) %	94±16 (91) %
Pancreas	86±21 (72) %	96±12 (96) %	87±20 (73) %	86±20 (72) %