EndoStyle: AI-based image style transfer for the optimization of computer-aided polyp detection systems in colonoscopy

Aims The use of computer-aided polyp detection systems has become a part of routine clinical practice. However, frequent false-positive detections have raised concerns for wider acceptance. We have developed an AI-based system named EndoStyle that is able to transfer the style of endoscopic images between different video processors, including image detail and color temperature. This study has two main objectives: first, to examine the ability of EndoStyle to adapt image styles across different video processors, and second, to evaluate whether a model trained on images generated by EndoStyle demonstrates improved performance.

Methods A StarGAN-v2-based image transfer model was trained on 239875 images from five different video processors. The model's effectiveness was tested with 20 endoscopists from 14 different centers, who viewed 28 ten-second colonoscopy sequences. They were subsequently shown three images and tasked to identify which were captured during the same colonoscopy sequence. The images came from different groups: the EndoStyle group, containing transformed images to resemble the processor used in the video sequence; the positive control group, with images from the same video sequence; and the negative control group, containing images from different processors. Parallel to this, two polyp detection systems based on YOLO11 were trained on images from public datasets (Pentax, Olympus and Fujifilm). The Baseline system was trained conventionally without preprocessing, while the Augmented system was supplemented with additional images generated by EndoStyle, converted to match the target processor (Olympus). Both models were tested frame-by-frame on 48 complete colonoscopy videos, comparing their sensitivity and specificity for detecting 43 polyps.

Results Endoscopists selected images from the positive control group, negative control group, and EndoStyle group in 88.47%, 12.29%, and 86.12% of cases, respectively. Both models, 'Baseline' and 'Augmented,' detected all polyps in at least one frame, with per-frame sensitivity rates of 63.18% and 57.26%, respectively (p-value=0.647). The Augmented model demonstrated a significant 8.3% reduction in frames with false-positive detections compared to the Baseline model.

Conclusions EndoStyle successfully adapts the visual characteristics of video processors, making these adjustments imperceptible to endoscopists eyes. Additionally, the use of synthetic data generated by EndoStyle enables a significant reduction in false-positive rates in computer-aided polyp detection systems by incorporating more specific images of the target processor in the training data. This image preprocessing technique could help in improving the quality of computer-aided polyp detection systems, and hence increase their clinical acceptance.

Publication History

Article published online:
27 March 2025