Quantifying Bowel Preparation in Wireless Capsule Endoscopy: Optimizing Latency with Structural Reparameterization in the WCEFastViT Model

Wireless Capsule Endoscopy (WCE) has revolutionized small intestinal disease diagnosis with its minimally invasive approach. However, its effectiveness greatly depends on image quality, influenced by bowel preparation, which is traditionally assessed subjectively by clinicians. To address this, the WCEFastViT model was developed in this study, aiming to objectively quantify bowel preparation quality in WCE. This model classifies images into four distinct quality levels and significantly accelerates the inference process through structural reparameterization. Achieving an ideal balance between latency and accuracy, it delivers classification accuracy of 92.34%, with remarkably low latencies of 2.80 milliseconds per frame on the GPU and 11.76 milliseconds on the CPU. Such advancements not only enhance the accuracy and speed of bowel preparation assessments but also offer clinicians a more objective and efficient diagnostic tool. Overall, the study demonstrates the potential of deep learning in improving medical image analysis and patient care in clinical settings.