Effects of Parallel and Distributed Learning on CNN Performance for Lung Disease Classification

The development of Deep Learning applications has sped up by using Graphical Parallel Units (GPUs), but even the finite capacity of a single GPU could be a limitation for the Deep Learning progress. To cope with this problem, it is possible to develop parallel training phases by distributing the workload between several GPUs in single and multiple compute nodes. Thus, distributed learning implies not only a reduction of the time requirements but also, slight accuracy variations that could have huge implications in high-precision areas like medicine. This paper analyzed the performance and accuracy effects of distributed and parallel learning of a Convolutional Neural Network (CNN) developed for X-ray disease classification. For this aim, the paper presents a complete framework for X-ray diagnosis. Furthermore, a new database was built from several public datasets. This database includes chest X-rays split into 5 classes (COVID-19, viral pneumonia, bacterial pneumonia, tuberculosis, and healthy). The experiments show a speed-up of the training phase of 1.7 using two GPUs, reaching a speed-up upper to 2 using three GPUs. The results prove that parallel training does not imply a negative effect on the accuracy results, showing even an increase in the final accuracy for the models trained with two or more GPUs, these results contribute to the feasibility of the use of distributed learning.