Deep Learning: An Empirical Study on Kimia Path24

Document Type

Conference Proceeding

Publication Date

6-20-2022

Publication Title

IEEE International IOT, Electronics and Mechatronics Conference (IEMTRONICS) Proceedings

DOI

10.1109/IEMTRONICS55184.2022.9795793

Abstract

Deep learning has a large interest in medical image analysis as studies have shown several machine learning algorithms were successful in predicting disease. However, more work in needed to better understand the batch size, epoch, and learning rates. An empirical study of image processing with deep learning was conducted on the KIMIA path24 dataset. The rotation, width shifting, height shifting shear range, horizontal flip, and fill mode was used. The network was trained and validated by a total of 22,591 images from the KIMIA path24 dataset. ReLU was used for the convolution layer and softmax for the fully connected layer. Results found the batch size is inversely proportional to the network accuracy, the accuracy of a deep learning network is directly proportional to the number of epochs it passes through, and the learning rate does not bring any change to the network. The network performs best within a preferred learning rate.

Comments

Georgia Southern University faculty member, Hayden Wimmer co-authored Deep Learning: An Empirical Study on Kimia Path24.

Share

COinS