電腦斷層深度學習影像在胸腹部血管鈣化的品質表現：假體研究

目的本研究評估電腦斷層掃描(CT)在胸部和腹部血管內鈣化影像品質的表現。特別是探討新興的深度學習重建法(deep learning reconstruction, DLR)影像品質，是否優於傳統的濾波反投影法(filtered back projection, FBP)以及目前廣泛應用的疊代重建法(iterative reconstruction, IR)。材料與方法使用兩廠牌(GE 和 Canon)具備產生 DLR 影像的電腦斷層掃描胸和腹部鈣化物假體。實驗使用具有不同濃度鈣化物的心臟鈣化假體(cardiac calcification insert, CCI)來模擬病人的血管內鈣化。將此 CCI 嵌入胸部(thorax phantom)或腹部假體(abdomen phantom)，以模擬人體胸和腹部。另外，採用外環假體(extension ring)包覆胸腹部假體，以模擬大體型的患者。胸腔的掃描有低劑量、常規模式、及雙能量模式。影像重組成 DLR、FBP 和 IR 三種影像。以雜訊(noise)、訊雜比(SNR)、對比雜訊比(CNR)代表影像品質，使用體積電腦斷層劑量指引(CTDIvol)代表掃描的劑量。以優值(FOMCNR)評估影像品質和劑量的綜合表現。另外，我們也以 Agatston score 及鈣化體積，將鈣化結果量化，並比較不同管電壓及重建法之差異。結果在影像品質方面，不論是胸腔或腹部，不同後處理的重建法均對雜訊產生差異。具體而言，FBP 呈現最大(差)的雜訊品質，而 IR 和 DLR 相對較佳。若加入劑量探討整體影像的綜合表現，DL 優於 IR 及 FBP。在以 120 kV 及 DLR 為基準的鈣化分數中，該模式下對於血管內非鈣化處的誤判最低。隨著管電壓的增加，鈣化分數呈現減少的趨勢，FBP 重建法的誤判分數最為顯著。此外，鈣化體積與分數趨勢相符，且在大體型方面誤差更為顯著。結論在電腦斷層掃描鈣化影像品質方面，掃描條件、重建方法及廠牌的不同均有顯著影響。針對血管鈣化的臨床 CT 掃描，建議使用自動管電流調控技術(ATCM)並採用 DLR重建法，以獲得更佳的影像品質和適當的劑量。

ObjectiveThis study evaluates the performance of computed tomography (CT) imaging quality in depicting calcifications within chest and abdominal vessels. Specifically, it examines whether the emerging deep learning reconstruction (DLR) technique surpasses the traditional filtered back projection (FBP) and the widely used iterative reconstruction (IR) methods in terms of image quality.Materials and MethodsCT scanners from two manufacturers (GE and Canon), capable of producing DLR images, were used to scan chest and abdominal calcification phantoms. The experiment utilized cardiac calcification insert (CCI) with varying concentrations of calcifications to simulate vascular calcification in patients. The CCI was embedded into thorax and abdomen phantoms to mimic human chest and abdominal structures. Additionally, an extension ring was used to envelop these phantoms, simulating patients with larger body types. Scanning modes included low-dose, routine, and dual-energy. Images were reconstructed using DLR, FBP, and IR methods. Image quality was assessed by noise, signal-to-noise ratio (SNR), and contrast-tonoise ratio (CNR), with the volume CT dose index (CTDIvol) representing the scan dose. The figure of merit based on CNR (FOMCNR) evaluated the overall performance of image quality and dosage. Additionally, the Agatston score and the calcification volume were also used to quantify the calcification results, comparing differences across various tube voltages and reconstruction methods.ResultsIn terms of image quality, different reconstruction methods significantly affected noise levels in both chest and abdominal scans. Specifically, FBP showed the highest noise levels, while IR and DLR were relatively better. When considering radiation dose, DLR outperformed IR and FBP in overall image performance. At a scan of 120 kV and DLR, this mode demonstrated the lowest misinterpretation in non-calcified areas of the vessels. With increasing tube voltage, the calcification scores tended to decrease, with FBP showing the most significant error in misinterpretation scores. Additionally, trends in calcification volume and scores were consistent, with more significant errors in larger body types.ConclusionIn CT imaging of calcifications, scanning conditions, reconstruction methods, and brand differences significantly influence image quality. For clinical CT scans of vascular calcifications, it is recommended to use automatic tube current modulation (ATCM) and the DLR reconstruction method to achieve better image quality and appropriate dose.