علي احمد رحماني

Ultrasound imaging is known as one of the newest imaging methods in the field of medicine. In recent years, the advantages of ultrasound imaging have made it faster, safer, and cheaper than other types of imaging in many cases. Identification of cancerous masses and their segmentation in medical images, especially ultrasound images, is a big challenge for correct diagnosis due to the high noise and low quality of the obtained images. This thesis aims to improve breast cancer mass segmentation in ultrasound images and deep learning algorithms are used as a basic method for segmentation in order to improve the final result. The used method uses two main parts of pre-processing and processing for the purpose of segmentation. In the pre-processing stage, data augmentation techniques such as rotation and zooming have been used in order to increase the number of dataset images with the aim of improving the performance of the model training process. Then, one of the basic deep learning networks called U_Net network, whose performance has been improved by adding blocks to increase resolution, has been used to segment the masses in the images. In the end, the results were compared and analyzed in terms of model accuracy criteria and overlap of similar areas and Dice. The obtained results are equal to 98.84% by the Accuracy criterion, 89.24% by the IoU criterion, and the result obtained by the Dice criterion is equal to 89.27%. In general, the results of the eva‎luation criteria showed that the innovations presented in the proposed method improved the performance of the deep learning model and favorable results were obtained for the segmentation of the masses in the images.

قطعه بندي تصاوير التراسوند , يادگيري عميق , قطعه بندي تومور پستان , شبكه U_Net

Ultrasound Image Segmentation , Deep Learning , segmentation of breast masses , U_Net

aliahmad rahmani

dr. ali asghar beheshti shirazi