SE_Resnet14: Design and development of deep learning architecture for kidney microscopy images grading

This research work focuses on detection of the grade of the cancerous kidney cells from microscopy images which depend on the experience of the specialist but there are times when the experts might not agree with their decision. There are ways in which a second opinion can be provided for the diagnosis of the image which is a computer-aided diagnosis, which helps to improve reliability of the experts. Automatic and accurate classification of kidney microscopy images is very important in the medical field to identify malignant tumors and to classify them into grades. There are advanced convolutions neural network (CNN) based methods which are yielding good results in classification of natural images, hence there are good opportunities for these grading techniques in the domain of biomedical computer vision and image processing. In this research, a hybrid CNN model has been designed, which includes SE (Squeeze-and-excitation) block, Resnet block, and a resnet14 algorithm to gather called as SEresnet14. The research has proposed, the combination of these algorithms and blocks to classify and predict the grade of the kidney microscopy images into five classes G0, G1, G2, G3, and G4. The results shows that the model achieves accuracy ranges from 90.66% to 93.81% for multi-class classification and 98.87% to 99.34% for binary classification and the accuracy of the model on the KMC dataset is calculated as 0.87, the overall f1 score is 0.86, the overall precision is 0.86, the overall recall is 0.87, and overall Jacard score is 0.77.