Sensitivity of Auto-Segmentation to CT Reconstruction Algorithms
Mittelstaedt D, Klahr P, Nejad-Davarani S, Schulz H, Goshen L, Morris E, and Glide-Hurst C. Sensitivity of auto-segmentation to CT reconstruction algorithms. Med Phys 2017; 44(6):3302.
Purpose: In effort to reduce imaging dose while maintaining image quality, low-dose computed tomography (CT) coupled with hybrid iterative reconstruction (HIR) or model-based iterative reconstruction (MBIR) is gaining momentum. However, it is unknown how their application may impact autosegmentation frequently used for treatment planning. This work explores the sensitivity and performance of a hybrid atlas and model-based segmentation algorithm in the prostate and head and neck (H&N) across several reconstruction algorithms. Methods: A sensitivity study was conducted by reconstructing raw CT sinograms for 5 prostate and H&N cancer cases with the following settings: standard filtered backprojection (FBP), HIR levels 1-6 (quantum mottle noise reductions of 0.89 to 0.55, respectively), and MBIR (Level 3, soft and standard filters). Reconstructed CTs were then inputted into previously validated auto-segmentation software and agreement was determined based on volume and Dice similarity coefficient (DSC), using FBP as the gold standard. Results: In regions with discrete attenuation differences (i.e. femur, sacrum, and pelvis bones for prostate and mandible and brain for H&N), HIR and MBIR segmentations agreed well with FBP (DSC > 0.98, volume differences < 2%). Bladder, eyes, brainstem, and parotid segmentations were generally insensitive to HIR and MBIR settings (DSC > 0.90). Smaller, complex H&N organs had increased variability with reconstruction algorithm including the optic nerves, (DSC 0.59-0.89, ∼30% larger than FBP) and pharyngeal constrictors (DSC 0.71-0.83, ∼20% smaller than FBP). In pelvis, patient-specific Results were obtained for the rectum (DSC 0.73- 0.90) and seminal vesicles (DSC 0.46-0.93), suggesting that additional modifications may be required. In small organs (i.e., lens and cochlea), auto-segmentation was not as reliable. Conclusion: Auto-segmentation was robust for HIR and MBIR for high contrast structures. Soft tissue organs are likely to need additional physician modifications to ensure accuracy. Future work will include benchmarking against physician contours and identifying when MBIR offers potential to improve auto-segmentation performance.