Real-time implementation of artificial intelligence algorithms for assisted diagnosis in COVID-19

Abstract

Background, motivation and objectives: Lung ultrasound has emerged as a promising technique for diagnosing and monitoring pneumonia, a critical complication of SARS-COV-2 infection. However, the lack of trained personnel in this field continues to limit its use. This work aims to extend the use of lung ultrasound by reducing the learning curve of junior technicians through computer-aided diagnosis. Real-time algorithms based on Artificial Intelligence (AI) have been implemented to guide the operator during exploration and suggest possible diagnoses based on lung artefacts found. Methods: To facilitate lung exploration and detect the presence of pneumonia, a real-time algorithm has been developed. It combines AI models implemented with Keras and Tensorflow 2 and signal processing algorithms using the Python language. The algorithm helps technicians obtain the best image conditions by guiding them through several coloured labels on the screen based on the region explored, movement and probe orientation, and similarity with previously labelled pulmonary images. The AI models first evaluate the acquired image to determine if it is suitable for processing. Once marked as valid, another AI model is used to detect the pleura, which is crucial to detecting typical lung patterns such as A-lines, pleura irregularity and B-lines through signal processing, and consolidations, the diagnosis of which is one of the main points of disagreement in the interpretation of lung ultrasound by physicians. Multiple parallel processes using an i7 octa-core CPU were necessary to achieve a continuous image refresh rate due to the computational cost of the AI algorithms. Results and conclusions: The algorithm achieved a processing rate of 16 frames per second without any delay, capable of detecting typical pneumonia artefacts in real-time. This is a sufficient image rate for the human eye to track fluently. In conclusion, real-time AI-based algorithms for pneumonia detection are possible using a CPU. Future work involves computational improvements to accelerate processing, such as using graphical processor units (GPUs) and novel Adaptive Compute Acceleration Platform (ACAP) devices.