Rotational invariant fractional derivative filters for lung tissue classification

Abstract

A new and powerful rotation invariant fractional derivative convolution neural network model is proposed for the classification of five categories of interstitial lung diseases. Fractional derivative convolution neural network model employs fractional derivative filters for the texture enhancement of the lung tissue patches instead of the raw image, is given as input directly to the convolution neural network. These FD filters are rotation invariant which solves the problem of rotation invariance of lung tissue patterns caused by pose variations of the patient during CT scanning. Also, the problem of the poor performance of most classifiers such as, support vector machine and K-nearest neighbours caused by an imbalanced dataset is solved, by oversampling the minority categories emphysema and ground glass patches, and under-sampling the majority category, micronodules patches. The experimental results are executed on the publicly available interstitial lung disease database which shows the fractional derivative convolution neural network model performs better than the state-of-art with average F-score and accuracy noted as 93.32% and 93.33% respectively.