Bu çalışmada, yapay US görüntüsü iki yeni artımsal yapay sinir ağı (ArÖz ve GArÖz) kullanılarak bölütlenmiştir. Öznitelik vektör elemanları, iki boyutlu ayrık kosinüs dönüşümü (2B-AKD) uygulanarak sadece alçak frekansları temsil eden katsayıların alınması (4x4 benek büyüklüğündeki pencereler) sonucunda oluşturulmuştur. Böylece hem öznitelik vektör boyutu azalmış hem de görüntüde var olan gürültülerin etkisi azaltılmıştır. Artımsal öz-düzenlemeli (ArÖz) ağın düğüm sayısı, eğitim sırasında ihtiyaca göre otomatik olarak belirlenmektedir. Ağın eğitiminden önce belirlenen eşikdeğeri $(E_D)$ düğüm sayısını kontrol etmektedir. Ağın en iyi düğümlerini belirlemek için sıkıştırma temelli bir yöntem geliştirilmiştir. Ağın eğitimi tamamlandıktan ve düğümler belirlenip etiketlendikten sonra, hem bölütleme işlemi hem de sıkıştırma işlemi paralel olarak gerçekleştirilmektedir. Ağ tarafından üretilen tüm düğümlerin ağırlıkları ve etiketleri kod sözcüğünü oluşturmak için kullanılmaktadır. Genetik algoritmalar ile eğitilen artımsal öz-düzenlemeli (GArÖz) ağ, US görüntülerindeki karmaşık doku dağılımını en iyi temsil edecek düğümleri bulmakta ve düğüm sayısını azaltmaktadır. Genetik havuzun derinliği 20 dizi olarak seçilmiş ve eğitim kümesinden rasgele alınan vektörler ile genetik havuz oluşturulmuştur. Ağın eğitiminde kullanılan uyumluluk fonksiyonu, aday düğüm ile temsil edilen vektörlerin hatasını en aza indirgeyerek yeni aday düğümün eğitim kümesinden en fazla vektörü temsil etmesine olanak sağlamaktadır. Her iki ağın yapay kist görüntüsünü bölütleme başarımları karşılaştırmalı olarak incelenmiştir. GArÖz ağının daha az düğüm kullanarak daha yüksek başarımlar verdiği gözlenmiştir.

Computer aided segmentation systems are being used in order to help physicians in diagnosis. Mainly, two operations must be implemented for computer aided segmentation; 1) feature extraction, 2) segmentation. Feature extraction is the process in order to determine the different tissue characteristics of the image. However, segmentation is performed for extraction of different (discriminating) tissues found in the image.In this study, two novel incremental artificial neural networks, incremental self-organizing map (ISOM) and incremental self-organizing map trained by genetic algorithms (GISOM), are proposed for the segmentation of tissues in ultrasound (US) images. The elements of the feature vectors are formed by using the discrete cosine transform (DCT). The performances of the ISOM and the GISOM network are investigated for segmentation of the phantom cyst image. Feature vectors are formed by applying 2D-DCT and then taking only the coefficients (windows of 4x4 pixels) that represent low frequencies. Thus, the vector size is decreased and also the effect of noise in the image becomes minimum. There are two critical parameters that must be taken into account in extraction of features from US images when DCT is used. First one is the size of the window and the other one is the number of coefficients that will be used in feature vector after transform. Selection of big window size increases the frequency resolution but decreases the time resolution. On the other hand, selection of small window size increases the time resolution whereas decreasing the frequency resolution. In such case, a paradox of satisfying time or the frequency resolution arises. Selection of window size depends on the structural characteristics of tissue like coarseness and repetition. Window size must be enlarged as tissue becomes coarser. However, if the window size is chosen very large, since the probability of occurrence of different tissues in the same window will be higher, identification of the tissue becomes more difficult. Furthermore, using large windows causes heavier computational costs so training times are getting longer. When the window size is very small, tissue characteristics are lost. Therefore, the selected window must have a size that is capable of identifying the tissue sufficiently. The ISOM is an incremental unsupervised neural network. The topology of the ISOM is not necessarily to be predefined. The topology and the number of nodes of the ISOM are automatically determined during the training. Only a threshold value (ED) needs to be estimated before the training of the network. The threshold value controls the number of classes (tissues) in the US images. Low threshold value reveals the details (different tissues) in the segmented image. ISOM network searches for the optimum nodes to discriminate between tissues in the feature space. A compression-based method that is developed for this purpose is firstly proposed in the study. Both segmentation and compression processes can be accomplished in parallel after the training of the network is completed, nodes are determined and labeled. The weights and labels of all the nodes generated by the network are used to form the codewords stored in the codebook. Codewords stored in the codebook are coded by using Huffman coding and sent to the receiver. The mean square error (MSE) value of the compressed image obtained by this method changes depending on the ED value. MSE values decreases when ED value is set to a low value, and increases for high ED values. It is observed that ISOM network produces high-performance segmentations for low values of MSE.In order to find the nodes that best represent complex tissue distribution in US images and to decrease the number of nodes, incremental self-organizing neural network trained by genetic algorithms (GISOM) is firstly proposed in this study. The weights of only the first layer nodes are searched by genetic algorithms. Nodes are determined during the training depending on need, and the optimum nodes are searched. Each node in the pool represents a unique different tissue. Fitness value calculation, reproduction, crossover and mutation operations are performed in each generation. The string that gives the best fitness value after 20 generations is assigned as the new node of the network. The fitness function used in the training of the network enables the new candidate node to represent the most number of vectors in the training set by minimizing the error of vectors represented by candidate node.The GISOM network gave the best performances with less number of nodes compared to the other network.