A Transfer Learning Based Approach for Lung Inflammation Detection

Pneumonia is a life-threatening lung inflammatory disease caused by viral or bacterial infection of the lungs. Early diagnosis is vital for efficient treatment. When detecting Lung Inflammation, CT (Computer Tomography) scans are generally analyzed. The CT screening of millions of people puts a huge burden on radiologists and manual detection may be faulty and tedious. Hence recent efforts are to come up with an approach or method to automatically detect risks of this disease with professional accuracy. Deep Learning excels in providing investigative and predictive analysis in the early diagnosis of pneumonia-like disease using regular radiographic images, indicating a pretty good accuracy in distinguishing between viral and bacterial pneumonia. In this paper, we applied the concepts of transfer learning in this task and have shown the performance of five different learning models to achieve a good result on a lesser number of datasets that it would get on a large dataset. As transfer learning reduces the burden of training of dataset from scratch and provides reusability, we apply this for our experiment. Additionally, we illustrate, using pre-trained architectures of ImageNet along with different optimizers, initialization techniques, dropout, and batch normalization, optimal learning can be obtained. Furthermore, we discuss how transfer learning can provide reasonably good performance for the augmented and preprocessed dataset.