Background: Gland is a prevalent organ in a human body, synthesizing hormones and other vital substances. Gland morphology is an important feature in diagnosing malignancy and assessing the tumor grade in colorectal adenocarcinomas. However, a good detection and segmentation of glands is required prior to the extraction of any morphological features. Objectives: The aim of this work is to generate a glandular map for a histopathological image containing glandular structures. The map indicates the likelihood of different image regions belonging to glandular structures. This information can then be used as a clue for initial detection of glands. Methods: A pipeline to generate the probability map consists of the following steps. First, a statistical region merging algorithm is employed to generate superpixels. Second, texture and color features are extracted from each superpixel. For texture features, we calculate the coefficients of scattering trans- form. This transformation produces features at different scale-spaces which are translation-invariant and Lipschitz stable to deformation. To summarize the relationship across different scale-spaces, a region-covariance descriptor, which is a symmetric positive definite (SPD) matrix, is calculated. We call this image descriptor, scattering SPD. For color features, we quantize colors in all training images to reduce the number of features and to reduce the effect of stain variation between different images. Color information is encoded by a normalized histogram. Finally, we train a decision tree classifier to recognize superpixels belonging to glandular and nonglandular structures, and assign the probability of a superpixel belonging to the glandular class. Results: We tested our algorithm on a benchmark dataset consisting of 72 images of Hematoxylin & Eosin (H&E) stained colon biopsy from 36 patients. The images were captured at 20× magnification and the expert annotation is provided. One third of the images were used for training and the remaining for testing. Pixels with a probability value greater than 0.5 were considered as the detected glands. Table 1 shows that, in terms of the Dice index, the proposed method performs 5% better than local binary patterns and the combination between scattering SPD and color histogram results in 25% better accuracy than the baseline. Table 1: Average Segmentation Performance ApproachesSensitivitySpecificityAccuracyDice?Farjam et al. (baseline)0.50 ± 0.130.80 ± 0.150.62 ± 0.090.59 ± 0.14 superpixels + local binary pattern0.77 ± 0.060.67 ± 0.100.73 ± 0.040.77 ± 0.05 superpixels + scattering SPD0.77 ± 0.070.85 ± 0.090.81 ± 0.060.82± 0.06 superpixels + color histogram0.74 ± 0.220.82 ± 0.170.77 ± 0.10 0.79 ± 0.10 superpixels + scattering SPD + color histogram0.78 ± 0.07 0.88 ± 0.070.82± 0.060.84 ± 0.06 Conclusions: We present a superpixel-based approach for glandular structure detection in colorectal cancer histology images. We also present a novel texture descriptor derived from the region covariance matrix of scattering coefficients. Our approach generates highly promising results for initial detection of glandular structures in colorectal cancer histology images.


Article metrics loading...

Loading full text...

Full text loading...

This is a required field
Please enter a valid email address
Approval was a Success
Invalid data
An Error Occurred
Approval was partially successful, following selected items could not be processed due to error