Assessment of texture measures susceptibility to noise in conventional and contrast enhanced computed tomography lung tumour images

TL;DR

This study evaluates the robustness of seven texture measurement methods in CT lung tumor images under noise, identifying the most stable features to improve automated diagnosis accuracy.

Contribution

It compares the susceptibility of various texture measures to noise in CT images, highlighting the most noise-resistant methods for lung tumor analysis.

Findings

Wavelet packet and fractal dimension measures are least affected by noise.

Gabor filter and WP provide stable performance across datasets.

Fewer features in ROI can reduce noise susceptibility.

Abstract

Noise is one of the major problems that hinder an effective texture analysis of disease in medical images, which may cause variability in the reported diagnosis. In this paper seven texture measurement methods (two wavelet, two model and three statistical based) were applied to investigate their susceptibility to subtle noise caused by acquisition and reconstruction deficiencies in computed tomography (CT) images. Features of lung tumours were extracted from two different conventional and contrast enhanced CT image data-sets under filtered and noisy conditions. When measuring the noise in the background open-air region of the analysed CT images, noise of Gaussian and Rayleigh distributions with varying mean and variance was encountered, and Fisher distance was used to differentiate between an original extracted lung tumour region of interest (ROI) with the filtered and noisy reconstructed versions. It was determined that the wavelet packet (WP) and fractal dimension measures were the least affected, while the Gaussian Markov random field, run-length and co-occurrence matrices were the most affected by noise. Depending on the selected ROI size, it was concluded that texture measures with fewer extracted features can decrease susceptibility to noise, with the WP and the Gabor filter having a stable performance in both filtered and noisy CT versions and for both data-sets. Knowing how robust each texture measure under noise presence is can assist physicians using an automated lung texture classification system in choosing the appropriate feature extraction algorithm for a more accurate diagnosis.