Development of Deep Learning Medical Imaging Model for Interpreting Pneumonia Diagnosis

Abstract

Abstract: Pneumonia is still one of the key public health problems, particularly in resource-poor areas like Kisii County in Kenya, where poor diagnostic equipment hinders early diagnosis. In this project, the intention was to develop and evaluate an explainable deep learning approach for the diagnosis of pneumonia using Chest X-rays. The use of the DenseNet-121 model was considered appropriate since it is efficient in terms of its connection structure and, thus, enables effective feature extraction without the problem of vanishing gradients. Pre-processing of data for better generalization involved contrast limited adaptive histogram equalization (CLAHE), normalization, and data augmentation. Results obtained from the experiment showed a good performance, with an internal accuracy of 97.5% and an external accuracy of 93.6%. To overcome the problem of deep learning models considered as "black boxes," the model adopted explainability techniques such as Gradient-weighted Class Activation Mapping (Grad-CAM) and SHapley Additive exPlanations (SHAP) which allowed for providing explanations not only in the form of visual heatmaps but also feature-based ones to enable physicians to comprehend predictions. According to feedback from practitioners in the field, the incorporation of explainability was positively regarded due to the gained confidence and trust, thereby showing the possibility for adoption of the solution in practice. In general, this study proves the value of applying explainable AI solutions to provide an accurate and reliable diagnostic tool in rural areas where radiologists are rare.