Real-time small bowel visualization quality assessment in wireless capsule endoscopy images using different lightweight embeddable models

Abstract

Wireless capsule endoscopy (WCE) captures huge number of images, but only a fraction are medically relevant. We propose automated real-time small bowel visualization quality (SBVQ) assessment to eliminate transmission of irrelevant frames. Our aim is to design lightweight color-based models for segmenting clean and contaminated regions with minimal parameters, short training, and fast inference, suitable for WCE hardware integration. Using the Kvasir Capsule endoscopy dataset, we constructed models based on distinctive color patterns of clean and contaminated regions. While different classifiers have been trained and evaluated, the k-nearest neighbors (KNNs), multilayer perceptron (MLP), and gradient-boosted machine (GBM) obtained superior performance (accuracy: 0.87 +/- 0.12, Dice similarity score (DSC): 0.87 +/- 0.15, intersection over union (IOU): 0.80 +/- 0.19). Logistic regression (LR) had the shortest training and inference times. Our models offer simplicity, compactness, and robustness, delivering satisfactory real-time performance. Evaluation on the SEE-AI project dataset confirms good generalization capabilities, demonstrating practical solutions for WCE image analysis.