A Minimum Distance Cluster based on Region Growing

A Minimum Distance Cluster based on Region Growing Method Presented by Kai Zhao School of Computer Science and Technology, Hefei Normal University, Hefei, China kzhao@aliyun. com

Content Background • • Virtual 3 D human in scientific simulations The medical images are usually so ambiguous. Existing segmentation method are unstable Combination of region growing algorithms Proposed method • • • The seed detection Distance to the cluster MDRG algorithm Experimental Results • • Segmentation results Quantitative evaluation Conclusion

Background Virtual 3 D human in scientific simulations Where and how to got the models Real person or medical Images

Background The medical images are usually so ambiguous • Overlapping grey levels between object and background. • Even a specialist could not exactly discriminate the interested regions. Existing segmentation method are unstable. • Noted algorithm is the famous FCM and its derivations. • Image thresholding segmentation based on minimizing the measures of Fuzziness or maximum entropy. Naturally they are more useful , but, the solo region could not be easily separated independently.

Background Combination of region growing algorithms • Region growing algorithms self-growing by its neighbors could solve that segmentation problem in dispersed areas. • Center seeds manual positioning. • Region Growing by a minimum distance • Based on the grey distribution in histogram and plotted regions. ROI

Proposed Method • • • Given the distance is defined as the grey Euclidean distance the region whose minimum grey is larger than ROI is defined as a region of right (ROR) and is smaller a region of left (ROL)

Proposed Method Step 1. Find the peaks of h(g) and associate it with grey as a vector G. ----h(g) denote the occurrence number of pixels at grey of g in the detected sub-region Otherwise go to. Step 3 Find the candidate pixels in the detected sub-region Xx which have the same grey as G. The candidate seed sets can be obtained by

Proposed Method Step 4 the ultimate seed is selected from S, where its neighbor number is maximal The seeds detected from above are aimed at avoiding the similarity between different areas, but just the difference in grey value is not enough. Distance to the cluster • The distribution of the grey based on the histogram and the grey variance on images are referred as ---h(g) on the histogram and the grey variance on images

Proposed Method • Minimum distance cluster based on region growing

Experimental Results l Segmentation results Fig. 3. The segmentation result

Experimental Results In order to be compared with TMMF(minimizing the measure of fuzziness of an image ), the seed grey of are divided into {g (Lx) ~g (Ix)} and {g (Ix) ~g (Rx)}. TMMF are taken to determine the minimal grey value (min-TMMF) in {g (Lx) ~g (Ix)} and (max-TMMF) in {g (Ix) ~g (Rx)}. By the MDRG segmentation, the grey range from the minimal grey (Min-MDRG) and the maximal grey (max-MDRG) are obtained. Through changing the plotted area of Ix, Rx and Lx at times, the grey range in different experiments can be illustrated in Fig 4. (left). The errors of this method are compared with that of TMMF , which havsegmentation by TMMF of the muscles is shown e not gone beyond 3. 84 % The thresholding in Fig 4. (middle) which cannot distinguish the edges effectively when compared with Fig. 3 (b). So the proposed MDRG method in this paper can show best performance on ROI segmentations Fig. 4 Comparisons of MDRG and TMMF. Left: grey value result by the two methods. Middle: the thresholding segmentation of muscles by TMMF. Right: The histogram of ROI segmentation of Fig. 3 (b).

Experimental Results • Quantitative evaluation is performed using volumetric overlap error (VOE) and relative volume difference (RVD). To use the two indexes, Fig. 3 is converted to binary images to compare with those produced by manual segmentation method. For example, the VOE and RVD of Fig. 3 (b) are 10. 57% and 1. 19% respectively, because manual segmentation inevitably includes subjectivity, artificial errors are sometimes tremendous. So the results are within acceptable limits and the MDRG algorithm proposed in this paper works well.

Conclusion Medical image segmentation is a crucial step in image processing field which is a hotspot of research. ROI in medical images are often ambiguous so that it brought many difficulties in the segmentation. Based on the variance of pixels and the seed detection methods, a minimal distance cluster is designed and combined with region growing methods. In this paper, adjacent regions of ROI are used for training the seeds growing. To test its effectiveness, the grey range of the segmentation is compared with the thresholding segmentation methods which are based on minimal segment entropy and fuzzy sets theory. It shows that the positions of the plotted region have little influences on the result of the segmentation. or practical application, minute adjustment of Xx that obeys the conditions of the minimal distance cluster makes the proposed algorithm perform better. Finally, quantitative evaluation indexes of VOE and RVD are calculated for the segmentation. The result of the experiments shows that our methods is a stable one that operates well on ROI segmentation. However, some ROIs are composed of many dispersed regions which need more plotted areas for seed detections and segmentation adjustments. Moreover, medical images are often in low grey difference, so to use this method properly relies on more complex work, the errors of the segmentation cannot be wholly eliminated. In the future the complete automation for performing the proposed methods is more challenging. Even so, the segmentation result of certain ROI can not only provide accurate data for medical diagnosis and make a treatment plan for radiation therapy, but also establish the foundation for three-dimensional applications.

Thank you!