| Abstrakt: |
Nowadays, kidney lesions are becoming one of the leading causes of death around the world. But, early detection of kidney lesions may reduce the severity that leads to mortality. Automatic segmentation and classification of kidney lesions are performed nowadays with the help of deep learning techniques. However, existing techniques of kidney lesion detection are inefficient because of blurring at the edges during segmentation, which affects the classification accuracy and also has an over fitting problem. They also require a large amount of training time and memory. Chances of survival for patients are reduced when kidney lesion types are misdiagnosed. To resolve these issues, we proposed semantic multiclass segmentation and classification (SMSC) of kidney lesions. Here, we introduced the integrated edge detection module with dilated residual UNet (IED-ResUNet) for effective segmentation and the Hopfield convolutional neural network (HCNN) for effective classification of kidney lesions. Initially, CT images are pre-processed using the multilevel brightness preserving technique for noise removal and image enhancement. Then the pre-processed image is segmented using IED-ResUNet in which blurring occurs during segmentation is avoided by using the edge detection module (EDM). EDM avoids blurring by giving importance to edge features during segmentation of the kidney lesion. Then the segmented image is accurately classified into one of the eleven classes using HCNN. Finally, performance of the IED-ResUNet and HCNN are compared with the most commonly used segmentation and classification networks in terms of accuracy, specificity, dice coefficient, and recall. The proposed SMSC method achieves high accuracy (99.60%) compared to other existing approaches. [ABSTRACT FROM AUTHOR] |
Full text from SpringerLink