بخش‌بندي خودكار بافت و تومور در تصاوير سي‌تي كليه با استفاده از روش Attention Based در يادگيري عميق

Kidney cancer is one of the ten most common cancers in humans. The incidence of kidney tumors, especially small and localized tumors that are often discovered by chance, is increasing. Identifying and segmenting benign and malignant tumors in medical images is often difficult; But in the end, most kidney tumors are diagnosed as malignant. Doctors mainly rely on CT images to diagnose the exact location of the tumor. Semantic segmentation can be a useful tool for quantitative description of tumors and kidneys. Segmenting images manually is a very time-consuming process; Therefore, an automatic or semi-automatic segmentation method is needed in order to increase the accuracy and speed in kidney tissue segmentation, which can be achieved with the help of deep learning. In this study, a two-dimensional U-Net based model and Attention Based module were used for kidney and kidney tumor segmentation. A set for training the network, which includes 200 cross-sectional CT images with kidney tumors and corresponding labels from the KITS19 challenge, has been used as the data of this research. 138 images were used for network training, 19 images for eva‎luation and 41 images as test data. After the pre-processing steps, a two-dimensional U-Net network was trained for kidney detection. After that, using kidney image extracted from U-Net network, a U-Net Attention based network was trained for tumor detection. After kidney and tumor segmentation for 41 test data, the mean Dice coefficient of kidney is 0.94, accuracy 0.95, specificity 0.95, IOU 0.90, recall 0.94 and sensitivity 0.94, and the results for tumor with mean Dice coefficient 0.79, precision 0.85, specificity 0.84, IOU 0.72, recall 0.80 and sensitivity 0.80 were obtained. The obtained results show that the proposed model has a good performance for kidney tissue and tumor segmentation.

5

حسين عربي