This study was conducted with the question of whether the use of AI can easily and accurately read SBCE images by removing poor visualization images. We identified that frame reduction reading after removing poor visualization images using the AI algorithm significantly reduced the reading time without affecting final SBCE diagnosis when compared to whole frame reading. Regarding the overall mucosal visualization score, the mean score difference between the AI measurement and the physician measurement for each case was 0.29 ± 0.88, showing no significant difference.
In this study, when poor visualization images were removed using AI for each case, a mean of 40.6% of total SBCE images were removed. Also, when removing poor visualization images, a mean of 53.4% of total lesion images were removed together. However, the diagnostic yield was maintained because there was no case in which the lesion was completely removed in this study. SBCE recorded images at a rate of 2 frames per second for 8 to 12 hours (3). Although it was difficult to detect SB lesions in poor visualization images, there were cases in which SB disease was diagnosed by detecting the lesions in other adequate visualization images. Especially for polypoid or ulcerative lesions, additional lesion images may be taken because the CE passage rate was reduced due to the lesion.
When comparing the AI non-user group and the AI user group, the removal rate of images according to lesion was 48.7% for ulcerative lesion, 53.7% for hemorrhagic lesion, 65.0% for polypoid lesion, 10.2% for vascular lesion, and 25.1% for nonspecific lesion. Most (112,892 images, 78.1%) of the 144,612 total lesion images removed were hemorrhagic lesion. Because active bleeding and/or a large amount of blood clots often covered the mucosa, the AI algorithm recognized them as poor visualization images, leading to their removal. Also, the removal rate of polypoid lesion images was higher than the mean removal rate (53.4%) of total lesion images. Polypoid lesions were often found in the distal small bowel, and in the case of distal small bowel, mucosal visualization was often poor. As a result, polypoid lesion images were thought to be removed more than other lesion images.
Because a mean of 40.6% of SBCE images were removed in each case, the reading time in frame reduction reading was also significantly shorter than in whole frame reading. In particular, this study was noteworthy in that both the AI non-user group and the AI user group were read using the same software viewer commercially available method. The recent trend of SBCE reading using AI algorithm is computer-assisted lesion detection. It is necessary to install the AI algorithm into the software viewer in order to use AI in clinical practice (12). In a previous study related to lesion detection of SBCE using AI algorithm, it was not possible to reproduce the same SBCE reading as in real clinical practice because the AI non-user group and AI user group were read in different ways (13). However, in this study, SBCE reading could be performed using the same software viewer even after AI preprocessing
It was also possible to identify the SBCE mucosal visualization score using an AI algorithm in our study. The mean difference between mucosal visualization scores measured by physicians and the AI algorithm was almost similar at 0.29 ± 0.88. The majority of the score differences between physicians and AI algorithm were less than one point. However, a difference of more than two points was confirmed in four cases. Unlike the AI, physicians judged all cases to be poorer visualization. In the AI non-user group, subjective judgment and inter-observer variation should be taken into account because the physician measured the visualization score simultaneously with the SBCE reading (14). The guideline recommended writing bowel preparation quality in the interpretation for performance measures of SBCE. It also recommended maintaining adequate bowel preparation in more than 95% of elective SBCE (5). The purpose of bowel preparation for SBCE was to read more accurately by reducing poor visualization images. However, there were still controversies about the optimal bowel preparation method, type, and time for SBCE (5) (15) (16) (17). According to a recent randomized controlled trial, performing bowel preparation before SBCE did not improve diagnostic yield and mucosal visualization compared to clear fluids only (18). Although manual and AI calculated methods have been proposed for measuring bowel preparation quality (5) (19), there was no clear consensus on the measurement of SB cleanliness either. Therefore, further studies using the AI algorithm are needed to investigate the effect of bowel preparation at full-length SBCE video level and the consistency of bowel preparation scores with expert endoscopists.
Although the results of our study were good, we considered that it was challenging to apply this frame reduction method alone to real clinical practice. Also, this study had several limitations. The most important limitation is that there is a possibility that AI can erase meaningful frames that contain some significant lesion (Fig. 3). The images of blood over the mucosa and significant lesions obscured by the bowel residues may be missed by this proposed method. Of course, in this study, there was no difference in diagnostic accuracy because lesions were found in another images, but mistakes and errors can occur with using the frame reduction reading method alone. In this study, in 3 cases, the diagnosis was not consistent between the two examiners. One polyp, one bleeding, and one angiodysplasia case were misdiagnosed in the AI user group. When AI removed poor visualization images, significant lesions were also removed, and the AI user group diagnosed these cases as a nonspecific finding. Therefore, frame reduction reading should be used as adjunct to AI algorithm integrated with lesion detection reading. For example, after performing lesion detection with AI, poor visualization images can be removed by frame reduction to provide sharp and clear images. Second, there were not many participated endoscopists and SBCE cases in our study because this study was conducted in a single center. The effectiveness of the frame reduction reading should be confirmed with a multi-center and large-scale study in the future. Second, poor visualization images were removed with AI, put back together, and then uploaded to the software viewer. Third, the MiroCam used in this study is not approved in some regions or has a small market share worldwide. Therefore, studies on other SBCEs will also be needed. software processing that can be performed simultaneously with AI filtering during SBCE reading is required.