Conclusions

From segmentation of the MR images, it can be observed that DA provides the best performance in terms of accuracy and stability among all discussed

(a) Lesion marked by (b) Segmentation affected by (c) k-means-segmented physician nonuniform image (four clusters)

illumination

(a) Lesion marked by (b) Segmentation affected by (c) k-means-segmented physician nonuniform image (four clusters)

illumination

(d) DA-segmented lesion (e) AFLC-segmented lesion (f) AFLC-segmented lesion (four clusters) (eleven clusters) (four clusters)

Figure 6.20: Segmentation of cervical lesion.

(d) DA-segmented lesion (e) AFLC-segmented lesion (f) AFLC-segmented lesion (four clusters) (eleven clusters) (four clusters)

Figure 6.20: Segmentation of cervical lesion.

clustering algorithms in several aspects. It is unsupervised; since theoretically it is designed to reach global minimum, the result is not biased by initialization. For the same type of image, the pseudotemperature reduction rate can be fixed, thus the segmentation process does not need parameter manipulation thereby yielding fully automated processing. Although the processing speed of DA is slower than other clustering algorithms, for small images such as the ones used above (217 x 181), the processing speed of DA is comparable to the other algorithms used. DA is also noise tolerant because of its statistical nature.

Both k-means and FCM, the well-known clustering algorithms, suffer from the initialization and local minimum problems. Cluster initialization is crucial in yielding satisfactory results. When not initialized properly, a clustering algorithm might be trapped in a local minimum, failing to proceed to the correct cluster. Our experimentations show that with random initialization, both k-means and FCM fail to generate the lesion clusters in MRI MS segmentation. AFLC is an automated and adaptive improvement over k-means and FCM by incorporating neural leader clustering and FCM. The performance is improved; however, similar problems are still encountered. Initialization is eliminated by selecting the first incoming sample as initial centroid, therefore, the outcome is sample-order dependent. DA is the best candidate for medical image segmentation by an advanced clustering technique. It is not sensitive to parameter tuning, and initialization problem, and is noise tolerant and guaranteed to converge.

Advanced clustering techniques can provide general solutions for effective segmentation of a broad range of medical images. All segmentation examples presented in section 6.3 use image intensity as the single feature to clustering algorithms to demonstrate the efficiency of the algorithms. In real applications, local property or connectivity of adjacent pixel can be embedded into segmentation to achieve more accurate segmentation [66, 67].

0 0

Post a comment