Abstract:
In low-resource healthcare environments, limited access to radiologists and a high
computational burden of conventional deep learning models are significant challenges for
pneumonia diagnosis. The current Convolutional Neural Networks (CNNs) for chest X
ray classification demand high memory, computation and specific hardware, which is not
feasible in under-resourced hospitals and clinics. In this work, the authors explored the
possibility of a lightweight shallow CNN producing reliable pneumonia classification
without being too computationally expensive. The architecture proposed was comprised
of three convolutional layers to reduce the amount of computational and memory
resources without compromising the diagnostic performance. The model was tested on a
set of common experimental settings, with the other popular lightweight and deeper CNN
models. To compare the performances fairly, all the models were optimized in the same
way, trained by the same augmentation methods and assessed by the same evaluation
metrics. It involved two data sets: the first was a secondary benchmark data set from
publicly available chest X-ray repositories, and the second was a Kenyan primary data set
collected from health care facilities. The accuracy, precision, recall, F1-score, and Type I
and Type II error rate were used as evaluation metrics. The proposed model was found to
be 91% accurate on the secondary benchmark data set and 95% accurate on the primary
data set of the Kenyan. The model showed better stability of convergence, reduced
overfitting and reduced majority-class bias in comparison with deeper architectures. The
study proposes a framework for CNN that is efficient in terms of computation and
scalable, which is suitable for resource-limited healthcare systems, where there are limited
GPUs, memory, and a lack of digital infrastructure in low-resource clinical environments.
Keywords: Pneumonia, Lightweight, Imaging, Resource-Constrained, Artificial
Intelligence.
