Development of Lung Disease Classification Technique

Abstract

Automating medical diagnosis classifying Lung tissues mapping them to their relevant diseases has been a challenging task despite numerous of research been conducted. We have targeted four major lung symptoms, they are Micronodules, Ground Glass, Fibrosis and Emphysema. Healthy Lung is also taken into account to achieve optimized and distinctive multi-class classification. Feature Enhancement of disease patterns has been conducted and played a vital role in enhancing problem areas indicating lung disease later supporting the diagnosis. Feature selection targeting the discriminating features of each disease itself is a challenging task and has been done with the intentions of exploiting dominant structures of lung patterns. Our proposed feature enhancement technique extract coarse and fine, high and low, horizontally directed and vertically directed, lastly line like and non-line like components of an image. Our proposed Feature descriptor includes known Texture descriptors like Local Binary Pattern, Rotation Invariant Local Binary Pattern, Block Difference of Probabilities and Bilateral Filters, Gradient feature descriptors include Gradient Central Differential , Sparse Distributed Localized Gradient Fused Features of Objects and Spatial Stimuli Gradient Sketch Model, lastly there are gradient Orientation descriptors used which is Histogram of Oriented Gradient and Gradient Orientations. Feature Descriptor is then used for classification by inducing full dimension vector set which is quite large and another time dimension of feature vector is reduced using Eigenvalue and Correlation between feature vectors as Intrinsic Dimension Estimators and Local Linear Embedding,Principal Component Analysis Laplacian eigenmaps and Kernel Local Discriminant Analysis as Dimension Reduction technique before letting into a classifier. Latest classifier Extreme Learning Machine with its two forms kernel and non-kernel based is used for classification.We therefore successfully achieved 100% accuracy. Support vector machine is also used as a classifier for comparing overall results.