脈衝耦合類神經網路於磁振影像之腦組織分割

Brain Segmentation by Pulse Coupled Neural Networks in Magnetic Resonance Imaging

傳家啓(Jachih J. C. Fu) ； 蔡志文(Jyh-Wen Chai) ； 陳啓昌(Clayton Chi-Chang Chen) ； 王安祥(An-Hsiang Wang) ； 盧宗斌(Tzung-Bin Lu) ； 徐賢鈞(Hsien-Chun Hsu)

《中華放射線醫學雜誌》 30卷2期 (2005/05) Pp. 83-94

摘要

腦部磁振造影以二維影像格式擷取三維空間之資訊，如需以三維視覺化方式顯示大腦組織結構，首先需自二維磁振造影影像中切割出大腦組織結構，再經表面描繪或體積描繪方式重建切割出之大腦組織結構。因腦部磁振造影所產生的龐大資訊量，以手動方式執行大腦組織結構影像切割為一臨床上不可行之工作，因此一有效之大腦組織分割演算法為大腦組織結構三維視覺化及影像融合不可或缺之工具。本論文結合脈衝耦合神經網路(Pulse Coupled Neural Network, PCNN)及期望最大化(Expectation Maximum, EM)演算法自動化分割腦部磁振影像之白質(WM)、灰質(GM)及脊髓液(CSF)等組織。以模擬影像加入雜訊與非均一性兩因素為測試影像，Bias Corrected Fuzzy C-Mean (BCFCM)演算法為對照組，Jaccard Similarity為績效衡量指標。實驗結果顯示，非均一性因素未顯著影響分割結果，整體而言，EM-based PCNN演算法之分割績效顯著優於BCFCM ，然而於高度雜訊之情況下，EM-based PCNN與BCFCM兩者之分割績效，均無法穩定且顯著地淩駕對方。以本論文所提EM-based PCNN分割之大腦皮質、融合頭顱邊界檢測等輸出，可提供三維視覺化之影像融合效果。

關鍵字

腦部磁振造影 ； 期望值最大化 ； 影像處理 ； 脈衝耦合神經網路

並列摘要

Image segmentation for brain Magnetic Resonance Imaging is essential to visualize the 3D structure of brain. Since brain MR generates a huge number of images, manual segmentation is impractical In this paper, pulse coupled neural networks are combined with the expectation maximum algorithm (the EM-based PCNN) to automatically segment GM, WM and CSF in brain MR images. The segmentation performance of the proposed algorithm is compared with the Bias Corrected Fuzzy C-Mean (BCFCM) from the simulated image database, with six noise levels and two nonuniformity levels. The Jaccard similarity is used as the criteria of performance eveluation. Data analysis showed that the nonuniformity was not a significant factor. Overall, the EM-based PCNN outperformed the BCFCM. However, the EM-based PCNN is not able to maintain it advantage over BCFCM consistently under the highly noised condition. To visualize the 3D effect, this paper demonstrates a 3D visualization example by fusing the EM-based PCNN brain segmentation and the head contour border.

並列關鍵字

Brain ； MR ； Expectation maximization ； Images ； processing ； Pulse coupled neural network