Classification of parenchymal abnormality in scleroderma lung using a novel approach to denoise images collected via a multicenter study

Abstract

Rationale and objectives: Computerized classification techniques have been developed to offer accurate and robust pattern recognition in interstitial lung disease using texture features. However, these techniques still present challenges when analyzing computed tomographic (CT) image data from multiprotocols because of disparate acquisition protocols or from standardized, multicenter clinical trials because of noise variability. Our objective is to investigate the utility of denoising thin section CT image data to improve the classification of scleroderma disease patterns. The patterns are lung fibrosis (LF), groundglass (GG), honeycomb (HC), or normal lung (NL) within small regions of interest (ROIs).

Methods: High-resolution CT images were scanned in a multicenter clinical trial for the Scleroderma Lung Study. A thoracic radiologist contoured a training set (38 patients) consisting of 148 ROIs with 46 LF, 85 GG, 4 HC, and 13 NL patterns and contoured a test set (33 new patients) consisting of 132 ROIs with 44 LF, 72 GG, 4 HC, and 12 NL patterns. The corresponding CT slices of a contoured ROI were denoised using Aujol's mathematic partial differential equation algorithm. The algorithm's noise parameter was estimated as the standard deviation of grey-level signal (in Hounsfield units) in a homogeneous, non-lung region: the aorta. Within each contoured ROI, every pixel within a 4 x 4 neighborhood was sampled (4 x 4 grid sampling). All sampled pixels from a contoured ROI were assumed to be the same disease pattern as labeled by the radiologist. 5,690 pixels (3,009 LF, 1,994 GG, 348 HC, and 339 NL) and 5,045 pixels (2,665 LF, 1,753 GG, 291 HC, and 336 NL) were sampled in training and test sets, respectively. Next, 58 texture features from the original and denoised image were calculated for each pixel. Using a multinomial logistic model, subsets of features (one from original and another from denoised images) were selected to classify disease patterns. Finally, pixels were classified into disease patterns using a support vector machine procedure.

Results: From the training set, multinomial logistic model selected 45 features from the original images and 38 features from denoised images to classify disease patterns. Using the test set, the overall pixel classification rate by SVM increased from 87.8% to 89.5% with denoising. The specific classification rates (original/denoised) were 96.3/96.4% for LF, 88.8/89.4% for GG, 21.3/28.9% for HC, and 73.5/88.4% for NL. Denoising significantly improved the NL and overall classification rates (P = .037 and P = .047 respectively) at ROI level.

Conclusions: Analyzing multicenter data using a denoising approach led to more parsimonious classification models with increasing accuracy. This approach offers a novel alternate classification strategy for heterogeneous technical and disease components. Furthermore, the model offers the potential to discriminate the multiple patterns of scleroderma disease correctly.

Figure 1

Noise characteristics in a CT image. (a) Original CT image and histogram from background and aorta. (b) Denoised image and frequency histogram of the noise image from decomposed algorithm [15] assuming that the noise in CT is white noise. (c) Denoised image and frequency histogram of the noise image from the decomposed algorithm [15, 17] assuming that the noise in CT is non-white noise.

Figure 2

Scatter plot of texture feature, standard deviation (std dev) on various dose level by different kernel from (a) original image and (b) denoised image. The dose level ranged from 8 mAs to 1024 mAs. The kernel of B10, B30, B50, B70, and B80 stand for smooth, standard, sharp, very sharp, and over-enhanced kernel, respectively.

Figure 3

Disease pattern and normal lung tissue in thin-section CT scans of the chest. An enlarged view of contoured ROI is shown on the left side of each image. (a) Lung Fibrosis (LF): Destruction of the lung parenchyma of reticular opacification, traction bronchiectasis and bronchiolectasis with increasing attenuation [16]. (b) Pure Groundglass (i.e. Groundglass without distortion): Destruction of the lung parenchyma moderately increased attenuation homogeneously in the absence of LF [16]. (c) Groundglass with adjacent LF: Groundglass that is located adjacent to LF. (d) Honeycomb (HC): clustered air-filled cysts with dense walls [16] (e) Normal lung tissue (NL): The mean CT attenuation value of the lung parenchyma is lower than that of disease cases.

Figure 4

Comparison between original and denoised images from Figure 3.

Figure 5

Pixel classification rate in test set of features from original and denoised images by disease pattern and all type. “org” and “deno” represent texture features from original and denoise image respectively.

Figure 6

Mean ROI classification rate from test set of features from original and denoised images by disease pattern and all type. “org” and “deno” represent texture features from original and denoise image respectively. Error bars represent standard error of upper bound confidence limits. There was significant differences in NL (*: p=0.037) and overall (**: p=0.047)

Figure 7

Comparison of Pixel SVM Classifications by original and denoised features using the ROIs from Figure 3. Legend: LF , GG , HC , NL . Note that image sizes were rescaled subjectively. Each dot represented the same size of pixel, which was sampled by rule of one out of 4-by-4 neighboring pixels.