بخشبندي خودكار ريه در تصاوير برشنگاري رايانهاي با استفاده از يادگيري عميق

Image segmentation is one ofthe most substantial subbranches of image processing and machine vision, which is performed in the form of basic and automatic methods. Diagnosing tumors, calculating the volume of tumors, identifying normal parts of tissue compared to its abnormal parts and drawing contour of tissues for treatment planning are among the most important applications of image segmentation in medical fields. Usually, in different radiation therapy methods, tissue contour drawing is done manually by the doctor using CT images, which requires high precision and doctor’s experience and is a time-consuming method. In recent years, with advancement of computers and use of artificial intelligence capabilities, automatic image segmentation methods based on deep learning are used. Automated methods are repeatable and require less interaction and time with the user. Lack of interpretability, requiring quality data for training and over fitting are among the most important challenges of automatic segmentation. In this research, segmentation of lung CT images has been done using deep learning method and U-Net architecture. The dataset used in this research is a collection of the kaggle dataset, which includes 17000 lung CT images in size 512×512 along with the corresponding mask that 900 images and masks have been selected for training and 110 images for model eva‎luation. Accuracy, DSC and IOU have been used in eva‎luation of this model. The average value of accuracy, DSC and IOU is respectivly 0.9906 ± 0.005, 0.9373 ± 0.05, 0.873 ± 0.11 and the loss value of this model is 0.02 that indicates the proposed model in this research has an acceptable performance.

4