ABSTRACT
With the evolution of the human race, the associated diseases have also evolved. Pneumonia treated as the simple flu and allergy in the early stages of its inception is now threatening to humankind in various shapes like SARs and Covid. The advanced disease requires equal treatments and diagnosis. Our research tried to find and classify pneumonia inflammation within chest x-rays (CXR) with very limited datasets and has attempted to ensure a global perspective, i.e. one that addresses all possible inflammation regions within the CXR. In addition to having medical grade classification outputs in terms of accuracy and recall, we have also guaranteed to meet the medical requirements of classification justification with the help of modified class activation maps (mCAM). The training of a model having a global perspective understanding is carried out with the help of capsules network cluster (CsNC), which enables us to learn various geometrical, orientation, and positional views of the inflammation within the CXR. Our 16-capsules cluster helped understand different views easily within the same CXR without going through any image augmentation, as generally required by current detection models, thus reducing the overall training and evaluation time. We performed extensive experiments on the RSNA pneumonia dataset of CXR images using a set of evaluation metrics. We have been able to acquire up to 98.3% accuracy with a 99.5% recall during our final trials. We tested our final trained model over generic x-ray images acquired from clinics and found promising results over that.