Design and implementation of differentiated analytics workflow for imaging diagnostics on the intelligent integrated digital platform InSilicoKDD

In this paper we have proposed a conceptual model of differential analytical scientific workflow for imaging diagnostics of abnormal pneumonias. The model is based on the method of deep convolutional neural networks (CNN) and the approach of Gradient-weighted Class Activation Mapping (Grad-CAM), which uses specific for the class information from the gradient, incoming in the last convolutional layer of CNN, in order to create rough map of localization of important region within the medical image. The proposed conceptual model is implement in Python with Google’s open source framework Tensorflow and Keras. In the development process, we have used customized Jupiter Notebook – Colab. For the creation and visualization plots we used Matplotlib. Neural Networks are pre-trained with ImageNet dataset, which consist of large amount of non-medical images. The experimental results for the 5 types of convolutional neural networks show accuracy more than 95%. The models for image diagnostics are integrated within the machine learning section of the intelligent platform for big biomedical data analytics InSilicoKDD.