Impact of ct model-based iterative reconstruction on auto-segmentation of prostate cancer organs at risk.
Miller C, Mittelstaedt D, Black N, Klahr P, Nejad-Davarani SP, Schulz H, Goshen L, Morris E, and Glide-Hurst C. Impact of ct model-based iterative reconstruction on auto-segmentation of prostate cancer organs at risk. Med Phys 2018; 45(6):e622.
Purpose: Model-based iterative reconstruction (MBIR) reduces imaging dose while maintaining image quality. However, it is unknown how MBIR may impact intensity-based tasks, such as segmentation. This work evaluates the sensitivity of an auto-contouring atlas in the prostate across multiple MBIR reconstruction filters and benchmarks the results against traditional filtered back projection (FBP). Methods: Raw projection data for 11 prostate cancer cases were reconstructed using FBP and MBIR at levels 1-3 for the following filters: body routine, body sharp plus, and body soft tissue, yielding 10 reconstructions per patient. Five bony structures and 5 soft tissue organs (bladder, rectum, prostate, and seminal vesicles (SVs)) were segmented using an auto-segmentation pipeline that utilizes a global positioning registration, landmark localizations, and an organ-specific positioning atlas to render 3D binary masks for analysis. Comparisons were performed for segmentation based upon MBIR filter and level that determines the image “smoothing”. Performance was evaluated for volume percent difference (VPD) and Dice similarity coefficient (DSC), using FBP as the gold standard. Sensitivity of each organ to MBIR filter was assessed. Results: MBIR reconstructions agreed with FBP for structures including sacrum, femur, and pelvic bones (DSC ≥ 0.98, VPD < 2.0%). For bladder, segmentations were generally insensitive to MBIR (DSC ≥ 0.92, VPD < 9.0%) while other soft tissue structures had increased variability based on MBIR settings. The VPD and DSC for prostate were 5.82 ± 7.24% (range: 0-51.77%) and 0.92 ± 0.09, respectively. The VPD and DSC for the rectum was 7.38 ± 7.94% (range: 0-42.8%) and 0.91 ± 0.10, respectively. SVs demonstrated the worst performance with VPD and DSC of 10.19 ± 15.15% range (0.0-113.34%) and 0.87 ± 0.16, respectively. Body routine MBIR filter outperformed all filters with DSC > 0.95 in 70% of comparisons. Conclusion: Auto-segmentation for MBIR on high contrast structures was successful, although complex soft tissue organs such as SVs require manual edits. Future work may involve tuning organ-specific MBIR parameters to improve autosegmentation performance.