Cascaded Deep CNN Pipeline for Pneumonia Screening: An Explainable AI based Application

Pneumonia is one of the major health problems to be faced throughout the world and has always hit the most vulnerable groups of populations-children, older adults, and those with other health conditions. The medical diagnosis, often CXR-based, faces errors due to limitation in image quality, radiologists’ fatigue, and human mistakes. This paper proposes the cascaded deep CNN pipeline along with Explainable AI (XAI) techniques to effectively detect pneumonia from CXR images. In this regard, there is a dataset of 2,023 labelled CXR images that were collected from a diagnosis clinic in Pabna of Bangladesh. The images were pre-processed using different pre-processing techniques such as CLAHE and noise-reduction filters for improving the quality of the images. From these pre-processed images, several pre-trained CNN architectures were then applied in extracting features: ResNet50, Xception, and InceptionV3, which were furthered with LR and XGB. Precisely, the best performance could be observed in models like ResNet50 with LR, which yielded 97.86% average accuracy, and Xception with XGB, yielding up to 98.02%. With this in mind, the model used techniques such as Grad-CAM to visualize the region of interest in CXR images for better transparency and interpretability of the model. This study thereby shows that a combination of CNNs with suitable classifiers and explainable AI techniques is a trustworthy and computationally efficient solution for pneumonia screening. This framework will be extended for other pulmonary diseases very soon and then validated in real-time clinical settings.