Abstract
SIGNIFICANCE: The identification of gastrointestinal bleeding holds significant importance in wireless capsule endoscopy examinations, primarily because bleeding is the most prevalent anomaly within the gastrointestinal tract. Moreover, gastrointestinal bleeding serves as a crucial indicator or manifestation of various other gastrointestinal disorders, including ulcers, polyps, tumors, and Crohn's disease. Gastrointestinal bleeding may be classified into two categories: active bleeding, which refers to the presence of continuing bleeding, and inactive bleeding, which can potentially manifest in any region of the gastrointestinal system. Currently, medical professionals diagnose gastrointestinal bleeding mostly by examining complete wireless capsule endoscopy images. This approach is known to be demanding in terms of labor and time. AIM: This research used white-light images (WLIs) obtained from 100 patients using the PillCam™ SB 3 capsule endoscope to identify and label the areas of bleeding seen in the WLIs. APPROACH: A total of 152 photographs depicting bleeding and 182 images depicting non-bleeding were selected for analysis. In addition, hyperspectral imaging was used to transform WLI into hyperspectral images using spectral reconstruction through band selection. These images were then categorized into WLIs and hyperspectral images. The training set consisted of seven datasets, each including six spectra. These datasets were used to train the Visual Geometry Group-16 (VGG-16) model, which was developed using a convolutional neural network. Subsequently, the model was tested, and its diagnostic accuracy was assessed. RESULTS: The accuracy rates for the respective measures are 83.1%, 65.8%, 66.2%, 72.2%, 73.7%, and 88%. The respective precision values are 78.5%, 47.5%, 30.6%, 59.5%, 77.7%, and 80.2%. The recall rates for the relevant data points are 83.3%, 67.9%, 86%, 74.2%, 68.6%, and 92.4%. The initial dataset comprises an image captured under white-light conditions, whereas the final dataset is the most refined spectral picture data. CONCLUSIONS: The findings suggest that employing spectral imaging within the wavelength range of 405 to 415 nm can enhance the accuracy of detecting small intestinal bleeding.