امير طهماسبي

Cancer is the main cause of death and an important obstacle to increase life expectancy in the countries of the world. In the process of cancer diagnosis, histological images are known as references and the type and severity of cancer can be diagnosed by using them. These images have large dimensions and their review is time-consuming. By automating the examination of these images, the workload of pathologists can be reduced and human error can be reduced. The method presented in this study was based on the combination of the features of U-Net and Atrous Convolution methods; In this way, the presence of a jump connection causes the transfer of features from the encoder to the decoder. In the proposed Cross O-Net structure, in order to increase the details and according to the dilated convolution, which has the ability to extract general and local features from the image. Two encoder parts with different fields of view have been used, which makes it possible to extract features with different levels from the image. The decoder section also includes two sections; that each part uses the output of one of the encoders and the jump connection of the other, this causes features with a different field of view from the image to be placed in the feature map, which ultimately leads to the division of images with details It gets better. In order to investigate different image segmentation methods, 5 databases were used. The HTS database was specially photographed for spheroid segmentation in the microfluidic laboratory of Shahid Beheshti University, and the range of each spheroid was determined so that these images could be used in the training and eva‎luation process of the presented method. Also, 4 other databases have been used to ensure the correctness of image segmentation. Each of these 5 databases used in this study were collected in different conditions and were used for a specific application. Creating variety in the database ensures that the presented method for dividing images can be practical and how it functions in different fields. The obtained results show that cross-jump connections provide an acceptable performance. and lead to the division of images with more details than other methods. HTS database images were segmented with the Jaccard index with 0.9201 by Cross O-Net structure and ResNet backbone; which showed better performance than other methods that were eva‎luated. Also, in other databases, considering the complexity of the database, the proposed method performed well.

ناحيه بندي تصاوير , تصاوير بافت شناسي , تشخيص سرطان , يادگيري عميق

Image Segmentation , Histopathology Images , Cancer Diagnosis , Deep Learning

Amir Tahmasbi

Dr Aboozar Ghaffari