Multivariate Gaussian Bayes classifier with limited data for segmentation of clean and contaminated regions in the small bowel capsule endoscopy images

Abstract

A considerable amount of undesirable factors in the wireless capsule endoscopy (WCE) procedure hinder the proper visualization of the small bowel and take gastroenterologists more time to review. Objective quantitative assessment of different bowel preparation paradigms and saving the physician reviewing time motivated us to present an automatic low-cost statistical model for automatically segmenting of clean and contaminated regions in the WCE images. In the model construction phase, only 20 manually pixel-labeled images have been used from the normal and reduced mucosal view classes of the Kvasir capsule endoscopy dataset. In addition to calculating prior probability, two different probabilistic tri-variate Gaussian distribution models (GDMs) with unique mean vectors and covariance matrices have been fitted to the concatenated RGB color pixel intensity values of clean and contaminated regions separately. Applying the Bayes rule, the membership probability of every pixel of the input test image to each of the two classes is evaluated. The robustness has been evaluated using 5 trials; in each round, from the total number of 2000 randomly selected images, 20 and 1980 images have been used for model construction and evaluation modes, respectively. Our experimental results indicate that accuracy, precision, specificity, sensitivity, area under the receiver operating characteristic curve (AUROC), dice similarity coefficient (DSC), and intersection over union (IOU) are 0.89 ± 0.07, 0.91 ± 0.07, 0.73 ± 0.20, 0.90 ± 0.12, 0.92 ± 0.06, 0.92 ± 0.05 and 0.86 ± 0.09, respectively. The presented scheme is easy to deploy for objectively assessing small bowel cleansing score, comparing different bowel preparation paradigms, and decreasing the inspection time. The results from the SEE-AI project dataset and CECleanliness database proved that the proposed scheme has good adaptability.

Fig 1. Structure of the proposed clean and contaminated region segmentation method.

Fig 2. Typical example of WCE images with their corresponding binary GT masks.

First and second rows: original images and their manually annotated binary masks, respectively.

Fig 3. Demonstration of the distributions and model-fitting performance on the concatenated RGB pixel intensity values of (a): clean and (b): contaminated regions.

In the first, second, and third rows, the empirical histogram and the fitted Gaussian distribution models have been plotted in the R, G, and B color components, respectively (as their color indicates).

Fig 4. Typical images from three different datasets along with their GT and predicted mask by the GBC model.

Fig 5. The predicted black-and-white masks by the GBC model for some corrupted images subjected to different degradations.

Fig 6. The generated black and white segmentation mask along with the original image and their corresponding GT masks of some randomly chosen frames.

(a) Original images. (b) GT mask. (c) Output mask by the U-Net. (d) Predicted mask of Pix2Pix. (e) Outcome mask of the GBC.

Fig 7. Scatterplot demonstrates a comparison between the area of clean region prediction by the automated GBC segmentation model and the manually annotation for the test images.