Abstract:
In rural communities, access to reliable healthcare is a major challenge. The rural people depend on traditional remedies to tackle different ailments. Hence, acute inflammation of the respiratory tract caused by viruses and bacteria could lead to a condition called pneumonia. This condition always proves difficult to manage with traditional medicine. Hence, there is a need to introduce a technology-driven solution for early detection of pneumonia. Thus, in this research, a Deep Convolutional Neural Network enabled mobile app for detecting pneumonia was developed and its performance was elaborately evaluated. This research was executed in five stages, namely Data Acquisition, Model Training, Development of Mobile Application, and Performance Evaluation. In the Data Acquisition stage, chest x-ray image dataset was acquired from Kaggle which was then preprocessed and stratified in the Data Preparation stage. The preprocessed and raw dataset were used to train a group of TensorFlow models in the Model Training stage. The models were then deployed to the mobile app during the Development of Mobile Application stage. In the final stage, which is the Performance Evaluation and Validation stage, the performances of the trained models and the mobile app was determined, and were subsequently validated by a trained radiologist. From the analysis of the results obtained, Strata 3 outperformed both Strata 1 and Strata 2, with accuracies of 95% and 97%, respectively, for the Normal and Pneumonia classes. The developed mobile app
responded differently based on the data Strata used in training the TensorFlow model. The 2nd and 3rd Strata, respectively showed a response time of 93.3ms and 92.73ms. During the validation substage, an error rate of 13.3% was obtained for the 3rd Strata when validated by a trained radiologist while an error rate of 26.6% was obtained for both the 1st and 2nd Strata When validated by the mobile application.
