Deep Learning for Lung CXR Abnormality Detection

Abstract

The detection of abnormalities is a crucial step that requires an early diagnosis for effective management of respiratory diseases. These diseases if not detected at an early stage can prove to be life threatening. The most efficient way to diagnose such abnormalities is to make use of lung CXR (chest X-ray) images which captures the essential details of heart and lung related diseases and offers insights which can aid the clinical procedure. There are many methods and techniques that can not only prove to be helpful in disease detection and classification, but these methods also minimize the risk of inaccurate detection. Such methods include the use of CNN and DCNN techniques which captures the relevant features from CXR images and predict the disease based on these collected features. However, these methods are complex in nature and can be termed as black box methods. It is very difficult to prove their truthfulness as they can be very challenging for humans to interpret. To overcome this issue, this research presents a novel framework that performs medical image diagnosis based on CXR images, reports uncertainty of its predictions and presents an explanation for its results using XAI techniques. Our goal is to develop an interpretable and explainable Convolution Neural Network (CNN) model tailored specifically for abnormality detection in lung CXR images. Our proposed model successfully classifies the disease labels from NIH ChestX-ray14 dataset. The success of model's classification and its evaluation performance is evident with the results of evaluation metrics. The model is evaluated using metrics such as accuracy, AUC-score, TNR, FNR and FPR. Along with correct label prediction our proposed architecture also quantifies the uncertainty and confidence in predictions. These results instigate trustworthiness in our model's output as the model clearly shows for which predictions it is confident and for which it is uncertain. Aside from being interpretable our proposed architecture also offers visualization results for our model’s prediction. The selected XAI techniques show how our model came up with a certain prediction making it an effective solution for disease classification in CXR images. Overall, our framework is modular in nature. It allows application of any deep learning mechanism for image detection to be used with our uncertainty estimation method. Moreover, explanations can be generated for selected mechanism using desired XAI technique. The modular structure of our proposed framework allows flexibility in selection which makes it a versatile and innovative method.