基於腦部影像之強度特性，目前的影像分割方法仍然不足以分辨腦部病變與正常組 織之差異，因此傳統的分割方法，多配合採用多光譜MRI的方式來提高性能。本文旨 在提出一個自建構分群的腦部影像分割演算法（SCBIS 算法），本方法僅使用單光譜掃 描的MRI腦部影像，而可以找出病灶的區域。在一張灰階的腦部影像，產生病變的區 域應該與正常腦部組織的分佈是不一致的，因此計算彼此之間的分歧，便可以區分出病 變區域。在本論文之中，我們的方法被實現為一個灰階值分割方法，採用經過事先統計 的正常腦部組織影像，經由自建構規則的生成方法和模糊神經網路建模技術，而形成可 基於體素強度檢測病灶的SCBIS 演算法。經由實驗證明，所提出的方法以實際影像模擬 病灶從20％到80％的區域，再與正常組織相比，經由Sensitivity、 Specificity 及Similarity index 三種驗證指標比較，本文的方法與FCM 演算法比較。時間複雜度在影像強度信號 減少到20％時，僅須0.1292 秒即可分辨，而FCM 演算法則須4.1798 秒。實驗證明本 文之SCBIS 演算法對偵測分割MRI影像的病灶是可行的。

Nowadays, image segmentation algorithm was still insufficient to identify lesions from normal tissues owing to their intensity similarity, so traditional segmentation methods might employ the multispectral MRI modalities to raise the performance. This paper intended to propose a Self-constructing Clustering Brain Image Segmentation algorithm (SCBIS algorithm) to segment stroke infarct lesions of MRI brain image by using the single spectral scan image. In gray intensity image shown, the lesion's regions should be inconsistent with the normal brain tissues distribution, so calculated the differences between each other then the lesions should be identified. In this approach, our method was implemented as a gray intensity segmentation method, while the normal brain tissues were provided by prior tissues probability maps. The SCBIS algorithm was modeled by using self-constructing rule generation method and Neuro-fuzzy modeling technology to detect infarct lesions based by voxel-intensity and adopted the intensity similarity of adjacent voxels' gray intensity to cluster to several fuzzy clusters. To validate quantitatively and effectively our research was efficient and feasible, both size and location lesions were predefined to form simulated MRI lesions were used to calculate the performance of the proposed method. To test the images of the lesions intensity signal reduction from 20% to 80% compared with normal tissues showed that the results were practicable at similarity index bigger than 0.8. The sensitivity of this approach was 0.8523, 0.9427, 1.0000 and 0.9988, and they were better than the compared method (FCM algorithm). The time complexities of the presented method also were better than the FCM algorithm of whole procedure. The SCBIS algorithm needed only 0.1292 seconds at reducing the intensity signal to 20% when compared with normal tissues, and it was much better than 4.1798 seconds had been needed by the FCM algorithm. Experiments showed the robustness of the SCBIS algorithm on MRI images. We will attempt to adapt the SCBIS algorithm to multiple sclerosis lesions detection in the near future.