Performance investigation of several convolutional neural network models in healthcare systems





Book title

Academic Press


Hala Shaari
Jasmin Kevric
Muzafer Saračević
Nuredin Ahmed


97 - 112


The main objective of this chapter is to explain and investigate the various CNN (convolutional neural network) models that can be used in healthcare systems (specifically in brain tumor analysis). Deep learning adaption for noninvasive brain tumor prediction can help doctors and radiologists in taking better decisions and reducing human error. To assist radiologists and improve diagnosis using magnetic resonance imaging (MRI), many studies have been presented. This chapter examines the behavior and efficiency of several CNN models with and without using transfer learning by applying five different CNN models to MRI images of brain tumors. We explain our proposed CNN models: SimpleConv, VGG, ResNet, and present the experiments followed by discussion of the results. A novel simple CNN model has been proposed and trained from scratch. Then, by fine-tuning the last layers of VGG and ResNet models, transfer learning efficiency has been tested. Findings suggest that all the proposed CNN models have achieved overall high accuracy, especially the Resnet-50 model, which achieved the highest accuracy with 99%. In medical imaging applications such as image sharing, developing collaborative models, and training a deep learning tool of global type, blockchain technology has been classified as the fastest growing technology in this cluster. Hospitals might protect their patients’ privacy by sharing just the weights and gradients of pretrained models (such as the models examined in this work) by utilizing blockchain technology to disseminate these data across the hospitals.

Publisher's website