Classification of Peripheral Pulse Morphology Using Deep- Convolutional Neural Network Based Transfer Learning Approach and its Implementation on GUI

Abstract

Peripheral Pulse (PP) is high-pressure wave of blood passing through blood vessels in the extremities. PP is recorded by using Peripheral Pulse Analyser (PPA) instrument and clinicians are using pulse morphology/pattern for diagnosing various diseases. The pulse morphology is seen to vary with respect to time in different individuals, also varies in multiple disease conditions, which results in struggle to estimate waveform accurately. Previously, Bhabha Atomic Research Centre (BARC) experts had classified eight patterns of PP but not enough to diagnose the disease accurately. Therefore, need to identify more number of patterns in PP waveforms and classify them using Artificial Intelligence (AI) techniques accurately. In this study, seven new patterns of PP are identified and Graphical User Interface (GUI) is developed using python language for the classification of fifteen patterns. Developed GUI is automatically predicts the pattern number of selected region and also gives information about percentage of each pattern in entire PPA signal. Each PPA signal consist more percentage of two or three pulse morphologies in the entire data of subject file. Then based on percentage of morphologies information, doctors decide about diseases like Hypertension, Pulmonary Tuberculosis, Cancer, Diabetes, and Coronary Artery Disease (CAD). Around 9000 images of fifteen different patterns of pulse waves are used for training and testing of AI models. The images resulting from fifteen patterns are classified using VGG16, VGG19 and ResNet50 existing deep Convolutional Neural Network (CNN) models with transfer learning approach. To assess each deep learning model's performance, four parameters namely accuracy, precision, recall, F-score are computed. After testing the dataset, the accuracy of VGG16, VGG19, and ResNet50 are 90.78%, 92.33%, and 91%, respectively. VGG19 model has provided best result as compared to VGG16 and ResNet50 models.