A Quantification Technique of Air Trapping in Lungs using Stepwise Regression and Neural Network from End-inspiratory and End-Expiratory CT-images

This work studies on air trapping problem in lungs using supervised machine learning technique. Firstly, we aim to detect an abnormal signal from the expanding or shrinking ability between end-inspiration (EI) and end-expiration (EE) phases of a breath which is calculated by using the EE to EI ratio $\rm{(Ratio_{EE-to-EI})}$ from the CT image pairs. Secondly, the attenuation of Hounsfield Unit (HU) is determined to explain the abnormality of air trapping problem. If there is an abnormal signal, the $\Delta$HU will be changed to be approximately lower than 100 HU. The alignment of pair images is manually set at the carina as a centroid. After that, the preprocessing is to set the window width (WW) and window levels (WL) as 1000 HU and -700 HU respectively. Finally, all features are combined and Stepwise Regression is used to select relevant features based on their similarities which reduced the feature dimension from 20 to 10 contributes. A traditional Neural Network and the proposed method are compared to improve the performance of the predictive model. It is found that the proposed method yields the accuracy (ACC) at 85.8% with 0.74 of F
1
score and the false negative (FN) are also reduced from 11.2% to 9.7%.