Transfer learning with NASNet-Mobile for Pneumonia X-ray classification

Abstract

Pneumonia affects 7% of the population worldwide and results in about four million deaths worldwide. The mortality caused by pneumonia can be prevented, as the treatment is low-tech and low-cost, yet it often goes unrecognized. The chest X-ray is the most reliable diagnostic imaging technique for pneumonia. Yet, often it is not used for lack of trained diagnosticians. However, this can be overcome with deep learning computer-aided diagnostic technology, which is shown in this study as well is in previous research to be able to achieve high performance in detecting and classifying between healthy and pneumonia radio graph images. This study presents a comparison between a transfer learning model based on NASNet-Mobile and a custom custom convolutional neural network (CNN) topology. Transfer learning has enhanced the model performance with an average of 5% for accuracy and lowered the loss with 15%. The experiments point to the fact that with fine-tuning, transfer learning can greatly improve custom CNN models. These results are significant as building transfer learning models based on simpler models can be faster and cheaper to industrialize and can be a viable option for providing the needed computer-aided diagnostic support system for pneumonia detection in chest radio graphs.