Characteristics of bowl images predict the quality of colon cleansing before colonoscopy

Aims The aim of the study was to detect predictors of bowel preparation from the characteristics of toilet bowl images to evaluate bowel preparation before the procedure. This will help in finding innovative solutions to improve the colonic preparation during colonoscopy.

Methods We collected 565 bowl images from 100 patients after the second dose of a split-dose preparation (using a high volume while adhering to a low-fiber diet for three days) between April 2023 and October 2023. The Boston Bowel Preparation Scale (BBPS) is divided into two categories: Good (≥8) and Inadequate (<8). We classified the visual appearance (CVA) of the toilet bowl as 1) clear or cloudy with slight debris or 2) semi-clear watery stools with cloudy or thick particles. For each image, we improved accuracy by using image cropping, allowing us to extract numerous scores expressing different image characteristics using MATLAB. We then specified the corresponding mean scores of the image in RGB (Red, Green, Blue) and HSV (Hue, Saturation, Value).

Results A threshold was defined for each image feature using the area under the ROC curve (AUC). Each feature was categorized into four quartiles. The binary logistic regression showed a significant correlation between blue (threshold=68, p=0.013), saturation (threshold=0.64, p=0.007), the first quartile of hue (threshold≤0.073; p<001), and the last quartile of brightness (threshold≥0.527, p=0.002). Then, a binary logistic regression formula was established using the following formula: 1.676 x first quartile hue+1.296 x last quartile brightness+1.082 x saturation (threshold=0.64)+1.142 x blue (threshold=97.5) – 0.774. The AUROC of the new distribution was equal to 0.716 [0.636-0.796] and revealed a threshold=3. The classification based on regression with a cut-off of 3 (CRC3) (BBPS as dependant factor) showed that the last image before colonoscopy had a sensitivity=80%, specificity=76%, and accuracy=85%, whereas the previous images had a sensitivity, specificity, and accuracy of respectively, 80%, 57%, and 78%. All images had a sensitivity=80%, specificity=72%, and accuracy=76%. Similarly, CRC3 with de CVA as dependant factor revealed the last image before colonoscopy had a sensitivity=80%, specificity=44%, and accuracy=76%, whereas the previous images had a sensitivity, specificity, and accuracy of respectively, 83%, 30%, and 57%. All images had a sensitivity=82%, specificity=31%, and accuracy=60%.

Conclusions Blue, hue, saturation, and brightness of the toilet bowl image are the primary predictors of preparation quality. By using specific thresholds, we have established a two-step model to predict the quality of colon cleansing before performing the procedure. This method will allow the adjustment of the preparation strategy before undergoing colonoscopy to improve the final outcome of the procedure.

Publication History

Article published online:
15 April 2024