Massive training in artificial immune recognition algorithm for enhancement of lung CT scans

Abstract

We proposed a pixel-based machine learning algorithm in the training of artificial immune recognition system (AIRS) to detect lung lesions in two-dimensional computed tomography (CT) scans. AIRS is an immune based algorithm which inspired by several biological mechanisms in mammalian immune system such as mutation, clonal expansion and immune memory generation. The proposed framework implements the concept of pixel machine learning (PML) where no segmentation and features calculation are required in the pre-processing of pixels. Hounsfield (HU) values in the selected region of interest (ROI) in CT scan are used directly to form a large number of learning sub-regions for massive training process. By using raw data in training, the loss of pixel information during detection of abnormality on medical images can be avoided. There are two versions of the AIRS (AIRS1 and AIRS2) algorithms are involved in the experiments of comparing their performance in the classification of medical images. The main advantage of these AIRS algorithms is to remove surplus training data while remain only relevant features in the processing of large amount of data training. The validation of results based on visualization validation and quantitative comparison using Kullback Leibler Divergence (KLD) are introduced. In this research, the massive training AIRS (MTAIRS) algorithms have generated promising results in visualization for lesions enhancement and detection in CT scans.