Fig. 1. The first proposed automated technique based on region growing segmentation. Fig. 2. The second proposed automated technique based on cellular neural network segmentation. Fig. 3. Result of pre-processing step. (a) Original image, (b) Contrast enhanced image, (c) Median filtered image. By applying a 2-layer neural network, an appropriate threshold is obtained for each image. The input and target matrices for train- ing ANN consist of the extracted values related to the intensity his- togram features from ROIs (see Table 1) and the obtained threshold of each image for region growing. After training, the obtained neu- ral network is capable of generating a relevant threshold for seg- mentation. As such, each image is more accurately segmented applying its own generated threshold. Considering the fact that the mammograms do not have the same intensity contrast, allocation of a constant threshold to seg- ment the image by region growing leads to inaccurate segmenta- tion results. Consequently, an automatic method is needed to determine an appropriate threshold. To this end, a trained artificial neural network (ANN) is proposed. Intensity histogram features for training ANN to produce an appropriate threshold. Table 1 Fig. 5. Structure of cell in CNN. Fig. 4. Two-dimensional CNN with 4 x 4 cells. Fig. 6. Two selected images taken from DDSM database to train CNN: (a) original images, (b) manually segmented images. Fig. 7. Two selected images taken from MIAS database to train CNN: (a) original images, (b) manually segmented images. Fig. 8. Post-processing step: (a) original image, (b) segmented image, and (c) segmented image after pot-processing. Zernike moments and their physical meaning. Table 5 Fig. 10. Results of segmentation based on region growing (first proposed method) corresponding two images: (a) original images, (b) segmented images, and (c) targe images. Fig. 11. Results of segmentation based on CNN (second proposed method) corresponding two images: (a) original images, (b) segmented images, and (c) target image Architecture of applied ANNs in first proposed technique. Table 9 Segmentation results related to the second proposed technique. Table 8 Segmentation results related to the first proposed technique. Table 7 Classification specificity, sensitivity, and accuracy rate in proposed techniques. Table 11 Architecture of applies ANNs in second proposed technique. Table 10 the second proposed technique, other ANNs to classify mammo- grams in MIAS and DDSM databases are used (see Table 10). Fig. 12. ROC curves for proposed techniques on MIAS and DDSM databases. Fig. 13. Results of different classifiers using 10-fold cross validation. Comparison of existing methods. Table 12 function (RBF)) and different evolutionary algorithms (such as par- ticle swarm, harmony, ant colony, and tabu search) can be applied in determination of an adaptive threshold and CNN templates in the first and second methods. These can be considered as future work. 4. Conclusions Breast tumor segmentation is one of the most important and crucial stages in medical image processing and pattern recognition. In this work, two automated methods were presented based on the improvement of region growing and CNN segmentations to pro- duce an adaptive threshold and appropriate templates, respec- tively, in order to preserve tumor boundary information to diagnose benign and malignancy in mammograms. In the first method, intensity features and ANN were applied to generate an adaptive threshold in the region growing stage of segmentation process. In the second method, a learning algorithm was proposed using genetic algorithm to determine CNN templates in segmenta- tion of images (in DDSM and MIAS databases). The proposed meth- ods have improved the CNN and region growing methods in breast tumor segmentation. To evaluate the performance of the proposed segmentation methods, after extracting the intensity, shape, and texture features in order to characterize the segmented tumors, a genetic algorithm was applied to select the appropriate features. Finally, the tumor classification was performed using different classifiers such as MLP, random forest, naive Bayes, SVM, and KNN. As a result, MLP produced the best diagnosis performance in both proposed methods.
