Automated Axial Triradius Detection and Classification using Deep Learning: A Preliminary Study for Non- Invasive Cardiac Screening

The study of epidermal ridges on fingers and palm has long been examined in connection with genetic characteristics and other medical disorders, such as congenital heart disease. The axial triradius is one of the features of the dermatoglyphics that has gained a lot of interest because of its use as a non-invasive means of assessing health in the early years. In the traditional approach to dermatoglyphic analysis, the process involves manual examination and interpretation, which can be labor-intensive and challenging in large-scale screening scenarios, especially in resource-limited environments. In the present work an automated framework based on palm-print image and deep-learning techniques for axial triradius analysis is explored. Regions of the palms were detected by MediaPipe and the axial triradius was localized by a custom-trained object detection model of YOLOv5. A Convolutional Neural Network (CNN) using the VGG16 architecture was then used for pattern classification between normal and disease palm. The entire framework was wrapped into a web-based application made with flask and mysql for easier deployment. To develop the model and evaluate, 500 palm-print images were used. The YOLOv5 model demonstrated high detection accuracy of 90% or above for the localization of the target anatomical landmark (axial triradius) in the palm-print image for different palm-print samples. The overall accuracy of classification experiments using the VGG16 network was 65%, and its precision was between 0.64 and 0.66, its recall was 0.65, and its F1-scores ranged from 0.64 to 0.65. The results validate the feasibility of the proposed automatic localization of landmarks and deep-learning–based analysis for dermatoglyphic screening applications. The proposed framework provides a basis for the automatic analysis of palm prints using general public image and computing tools. Future enhancements will involve increasing dataset diversity, adding additional dermatoglyphic parameters, and using more sophisticated learning architectures to achieve better classification accuracy and to enable future clinical validation studies.