Data augmentation method for computer-aided diagnosis using specular reflection.

Colorectal cancer (CRC) is a significant global health challenge, emphasizing the importance of effective screening by applying methods like colonoscopy. While advanced imaging technologies, such as narrow-band imaging (NBI), allow real-time optical diagnosis of colon polyps, variations in endoscopist skills and unnecessary polypectomy underscore the need for artificial intelligence applications, particularly deep learning (DL) in computer-aided polyp detection and diagnosis (CADe and CADx). This study developed and investigated a data augmentation technique using specular reflection (SR) to enhance the robustness and performance of DL models tailored explicitly for CADx in colonoscopy. This SR augmentation method included SR generation and inpainting integrated into conventional augmentation techniques. We utilized two DL architectures: a convolutional neural network and a vision transformer. Stress tests, under varying data usage ratios using a dataset of 2,616 NBI images, revealed the robustness of SR augmentation, especially in scenarios with limited training data, highlighting its superiority over other methods. SR augmentation effectively improved model accuracy, particularly in scenarios with limited data, supporting its practical implementation in real-world colonoscopy environments. The findings emphasize the significance of domain-specific data augmentation techniques to support DL application in colonoscopy imaging for more reliable and accurate CADx systems for colon polyps. Supplementary Information: The online version contains supplementary material available at 10.1007/s13534-025-00533-0.