Evaluation of Computer-aided Detection for Identifying Missed Gastric Cancer After Endoscopic Submucosal Dissection.
CADe showed no significant improvement in sensitivity for detecting missed gastric cancers compared to endoscopists, and both methods had low sensitivity overall.
Where it sits
this study against the rest of the aod-9604 corpusSummary and findings
The study evaluated the efficacy of computer-aided detection (CADe) for identifying missed gastric cancers (MGCs) after endoscopic submucosal dissection (ESD) using 2324 endoscopic images. The per-lesion sensitivity for CADe was 15.0% compared to 13.8% for endoscopist detection. The findings indicate no significant difference in sensitivity between the two methods.
Abstract
<h4>Objectives</h4>Computer-aided detection (CADe) using deep learning is promising for reducing missed gastric cancers (MGCs) and supporting physicians in double-checking endoscopic images. We aimed to evaluate the CADe efficacy for MGCs after endoscopic submucosal dissection (ESD).<h4>Methods</h4>We collected 2324 endoscopic images, including 60 of MGCs detected during surveillance esophagogastroduodenoscopy within 18 months after initial ESD. The performance of the CADe system, developed for early GC detection using deep learning, was compared with that of 10 endoscopists in a detection study using collected images. Per-lesion sensitivity, per-image detection performance, and diagnostic time were compared.<h4>Results</h4>The per-lesion sensitivity for MGCs was 15.0% (9/60) and 13.8% (83/600) for CADe and endoscopist detection, respectively (<i>p</i> = 0.538). The respective per-image computer-aided and endoscopist performance sensitivity was 11.3% and 10.6% (<i>p</i> = 0.548), specificity 88.5% and 94.8% (<i>p</i> < 0.001), positive predictive value 4.4% and 8.3%, and negative predictive value 95.7% and 96.0%. The per-image CADe time was significantly shorter (0.03 s vs. 2.86 s, <i>p</i> < 0.001). CADe showed higher sensitivity in certain subgroups, although these findings should be interpreted cautiously given the small sample size.<h4>Conclusions</h4>No significant difference in sensitivity was observed between CADe and endoscopist detection for MGCs after ESD, and the absolute sensitivity remained low. Further improvements are needed before clinical implementation of CADe as a double-checking tool.<h4>Trial registration</h4>The authors have confirmed clinical trial registration is not needed for this submission.
Background
This paper addresses the challenge of missed gastric cancers (MGCs) after endoscopic submucosal dissection (ESD), a significant concern in gastroenterology. Prior studies have indicated that computer-aided detection (CADe) using deep learning may assist in identifying MGCs, but its efficacy compared to human endoscopists was not well established. This study is important as it evaluates the performance of CADe in a real-world setting, potentially influencing future diagnostic practices.
Methods
The study collected 2324 endoscopic images, including 60 MGCs detected during surveillance esophagogastroduodenoscopy within 18 months post-ESD. The performance of the CADe system was compared with that of 10 endoscopists in a detection study. Primary outcome measures included per-lesion sensitivity, per-image detection performance, and diagnostic time.
Results
The per-lesion sensitivity for MGCs was 15.0% (9/60) for CADe and 13.8% (83/600) for endoscopist detection, with a p-value of 0.538. The per-image sensitivity was 11.3% for CADe and 10.6% for endoscopist, p = 0.548. Specificity was 88.5% for CADe and 94.8% for endoscopist, p < 0.001. The positive predictive value was 4.4% for CADe and 8.3% for endoscopist, while the negative predictive value was 95.7% for CADe and 96.0% for endoscopist. The diagnostic time for CADe was significantly shorter at 0.03 seconds compared to 2.86 seconds for endoscopist detection, p < 0.001.
Interpretation
The study found no significant difference in sensitivity between CADe and endoscopist detection for MGCs, suggesting that while CADe may offer time efficiency, its clinical utility remains limited due to low sensitivity. The findings align with previous literature indicating that while CADe can assist in detection, it does not replace the need for skilled endoscopists. The small sample size and low absolute sensitivity are confounding factors that limit the conclusions drawn from this study.
Key findings
- Per-lesion sensitivity for MGCs was 15.0% (9/60) for CADe and 13.8% (83/600) for endoscopist detection, p = 0.538.
- Per-image sensitivity for CADe was 11.3% and for endoscopist was 10.6%, p = 0.548.
- Specificity for CADe was 88.5% and for endoscopist was 94.8%, p < 0.001.
- Positive predictive value was 4.4% for CADe and 8.3% for endoscopist.
- Negative predictive value was 95.7% for CADe and 96.0% for endoscopist.
- CADe time was significantly shorter at 0.03 s compared to 2.86 s for endoscopist detection, p < 0.001.
Limitations
- small sample size of 60 MGCs
- absolute sensitivity remained low
- no significant difference in sensitivity between CADe and endoscopist
- further improvements needed before clinical implementation