用寬頻磁振造影重建高解析度臺灣人腦圖譜

Abstract

大腦圖譜在神經醫學與影像診斷上具有重要地位，能幫助研究大腦結構、功能連結與病灶變化。為了提升台灣族群的影像研究品質，本研究以寬頻磁振造影技術建構0.5 mm與1mm等向性高解析度的台灣人腦模板，並比較兩種降噪方法對影像品質的影響。 本研究招募114位受測者，排除不符條件者後，最終納入105位健康成人，皆簽署知情同意書。影像使用西門子 MRI 機器搭配寬頻序列掃描，取得 0.5 mm 及 1 mm 解析度的 T1 加權影像。針對影像中存在的低SNR與Rician雜訊，分別使用BM4D與AONLM兩種降噪方法處理，並進行圖譜構建、融合與分析。計算 SNR、SSIM 與 MAD 三項指標評估降噪成效。 降噪前，兩種解析度的影像皆呈現 SNR 偏低的現象。經處理後，BM4D 顯著提升 SNR， AONLM則更能保留邊緣與細節。個別影像中BM4D的SNR較高；但在群體融合影像中，0.5 mm解析度下 AONLM 效果更佳，1 mm 下 BM4D 則較穩定。融合處理能進一步提升影像穩定性與降噪品質。AONLM 在 0.5 mm 解析度下融合前後皆表現良好。 不同解析度適合不同降噪策略：高解析影像中，過度平滑會抹去細節，AONLM 更能保留結構資訊；較低解析度中，BM4D的全域平滑能有效提升SNR。融合處理能提升演算法效果，使降噪更精準。 本研究成功建構台灣高解析腦模板，並證實 BM4D 與 AONLM 均可有效提升影像品質。建議依據影像解析度與研究目的調整降噪策略與處理順序，以取得最佳分析結果。0.5 mm 等向性影像在掃描時間與品質間取得良好平衡，顯示其具實用性與應用潛力。

Brain atlases play a critical role in neuroimaging and clinical diagnosis by revealing brain structure, functional connections, and lesion localization. To enhance brain research quality for the Taiwanese population, this study developed a high-resolution 1mm and 0.5 mm isotropic brain template using wideband MRI and evaluated the effects of two denoising methods. A total of 114 subjects were recruited, and 105 healthy adults were included after exclusions. All participants gave informed consent. T1-weighted brain images at 0.5 mm and 1 mm resolutions were acquired using a Siemens MRI scanner with wideband sequences. To address low SNR and Rician noise, BM4D and AONLM were applied. Images were fused and analyzed using metrics including SNR, SSIM, and MAD to evaluate denoising quality. Before denoising, both resolutions exhibited low SNR. BM4D significantly increased SNR, while AONLM better preserved edges and anatomical details. For individual images, BM4D performed better in SNR. In group-fused images, AONLM outperformed BM4D at 0.5 mm resolution, while BM4D was more effective at 1 mm. Fusion improved image stability and enhanced denoising results. AONLM showed consistent performance before and after fusion in high-resolution data. Denoising strategy should be resolution-specific: in high-resolution images, preserving fine structures is essential, favoring AONLM; in lower-resolution images, BM4D’s global smoothing better improves SNR. Fusion further enhances denoising effectiveness by providing more stable input data. This study successfully reconstructed a high-resolution Taiwanese brain template. Both BM4D and AONLM improved image quality. Choosing the appropriate method and processing sequence based on resolution and research objectives is essential for accurate analysis. The 0.5 mm isotropic images offer a good balance of scan time and quality, demonstrating strong potential for future applications.