Abstract Book

S279

ESTRO 37

Poster Viewing : Poster viewing 10: Images and analyses

PV-0527 Gray-level invariant Haralick texture features P. Brynolfsson 1 , T. Löfstedt 1 , T. Asklund 1 , T. Nyholm 1 , A. Garpebring 1 1 Umeå University, Radiation Sciences, Umeå, Sweden Purpose or Objective Texture analysis is a common tool in radiomic analyses. Haralick texture analysis involves quantizing an image to a fixed number of gray levels (GL), N. The optimal N depends on e.g. the size of the analyzed region of interest (ROI). Texture features will only be approximately constant for a small range of GLs, and may drastically change as N changes. This is problematic if we want optimal features from ROIs of very different sizes. Features that are invariant to the number of GLs enable statistical models to be (I): constructed with varying quantizations in the same data, or (II): applied to differently quantized data. The aim of this work was to develop features that are invariant to the number of quantization GLs. Material and Methods By redefining the gray level co-occurrence matrix, GLCM, as a discretized probability density function, it becomes asymptotically invariant to the quantization levels (Fig. 1a). The original and renormalized definitions of the GLCM and the Energy feature are shown in Fig. 1b. We validated the new features using logistic regression (LR) on two datasets. Dataset 1 contained 63 T1-weighted MRI volumes of the brain, and texture features were used to classify extracted ROIs as cerebellum or pre-frontal cortex (Fig. 1c-d). Dataset 2 contained images of 1518 colorectal cancer glands in 165 Hematoxylin and Eosin (H&E) stained slides from the Warwick-QU dataset (Sirinukunwattana et al., 2015). Texture features were used to classify each gland as benign or malignant (Fig. 1e-f). Each dataset was split into training (50%), validation (25%), and test (25%) sets. Variable selection from 20 features (Brynolfsson et al., 2017) were used in the analyses to optimize accuracy, with quantization between 8 and 256 GLs in steps of 8. (I): LR classifiers were trained on data with random GL quantizations within the training sets. 30 test were evaluated, and compared with a t-test. (II): LR classifiers were trained on each GL quantization. We evaluated the accuracy of each classifier on all quantizations. A Mann-Whitney U test was used to compare the results.

Results Fig. 2 exemplifies how four features vary with quantization GLs. Most original features values diverge with N, while the renormalized features approaches a limit as N increases. (I): The average accuracy of classifiers trained on data with varying quantization GLs were 0.77±0.07 and 0.96±0.01 (p<1e-21) for the brain data, and 0.68±0.04 and 0.80±0.01 (p<1e-21)for the gland data, for the original and renormalized features. (II): Heatmaps in Fig. 2 show the accuracy of all train/test pairs for the brain and gland data. The renormalized feature accuracies are significantly larger (p<1e-99) for both data sets.