Author : Zahra Khamesi 1
Date of Publication :30th June 2023
Abstract: In collaboration with CIUSSS du Centre-Sud-de-l’Ile-de-Montréal, this paper presents RADIA, a project combining several techniques of deep learning algorithm developed for detecting 14 classes in the thoracic chest X-ray, using the combination of 112,120 public images. According to the ML group at Stanford, this concept was built based on previous studies on ChexNet. Our training phase has been extended by integrating Frontal and Lateral views and use of algorithms to improve the images. A model has been implemented based on ConvNeXt-Large , with improvements. We present the result obtain with the metrics AUC, F1, G-mean to better define the behavior with a imbalanced dataset. The challenge does not end with the realization of a software tool based on the deep learning technics but it will continue with the development of a decision helper combining inference model and webtool for the radiologists and other healthcare teams.
Reference :