Restoration, Enhancement and Analysis of Lung Nodular Images for Prompt Detection of Abnormalities

Abstract

Detection of lung cancer in the Computed Tomography (CT) images when the lung nodules are in the sub-solid state (early stage) results in higher survival rate of the patients. Two Computer Aided Detection (CAD) systems for identifying the sub-solid nodules in lung CT images are developed as a part of this thesis. The first system adopts a pipeline approach which is carried-out in two phases. The first phase employs a series of algorithms for denoising, segmentation of region of interest and feature selection followed by a classification to separate nodules and nonnodules. In the second phase, Histogram of Gradients method is used to categorize the nodules identified in first phase as solid or sub-solid. Sensitivity of the system is observed to be more than 90% with just 3 false positive observations per scan. Both supervised and unsupervised classification models adopted for identifying sub-solid nodules give consistent and reliable results with an average accuracy above 93% when tested with Lung Image Database Consortium (LIDC) and International - Early Lung Cancer Action Program (I-ELCAP) databases. The accuracy of the system is categorically higher compared to the present state-of-the-art models employed for sub-solid nodule classification. The second system adopts a deep learning approach for identifying sub-solid nodules, making use of a Deep Convolution Neural Network (DCNN) incorporated within the Conditional Random Field (CRF) framework. Adopting CRF framework reduces the occurrence of false positives. It is further observed that the overall accuracy of the system is increased from 83 to 89.5 percentage when tested with LIDC/IDRI and I-ELCAP databases. Though, the accuracy of the system is lower than the pipeline based model (mentioned above), the model does not demand any pre or post processing technique including the region of interest segmentation. The accuracy obtained for this system is comparatively higher than the state of the art deep learning models employed for sub-solid nodule classification. Moreover, a detailed cross comparative analysis of the systems proposed in this thesis is done to analyze their performance.