Classification of Chest X-ray images using radiomic features and machine learning

Abstract

Accurate identification of the existence of disease in a radiograph by a radiologist is highly essential. Various systems are being developed to assist radiologists to diagnose the disease with the best accuracy possible. Machine Learning algorithms have been used in classification tasks, particularly Convolutional Neural Networks, a variant of Neural Networks has been proven to outperform most algorithms. The success is attributed to this variant because they resort to optimal feature construction on their own instead of depending on a finite set of feature candidates. But CNNs in their basic form or various pre-built models like VGG Net, MobileNet, etc. are able to classify the images based on certain conditions i.e. have sufficient training data and expensive computing power. The classification of radiographic images for medical diagnosis can also be achieved usinga predefined set of features which are calculated via the extraction of quantitative metrics. Such a process is known as radiomics or radiogenomics (kumar et al, 2012).
We have extracted 955 features (shape, texture, transform) from the X-ray images and have identified the features that are very effective in the classification of normal vs. Pneumonia. Among such features are wavelet transform features particularly low pass (LH) gray level uniformity, Correlation of GLCM, Inverse Difference Normalized, High Gray Level Zone Emphasis, Gray Level Non-Uniformity Normalized, high pass (HH) gray level uniformity, low pass (LH) run length uniformity, high pass (HH) run length gray level uniformity. We were able to achieve accuracy of 99% for our primary dataset and 90% for the secondary dataset which is unseen by the model in training phase. In this paper we particularly discussed and investigated the dataset open sourced by (Khuzani et al, 2021) in a paper submitted to Nature journal. We get the same accuracy levels whether we use all the features that were calculated after applying wavelet transform or the few above mentioned features.