Quality controlling in capsule gastroduodenoscopy with less annotation via self-supervised learning.

BACKGROUND

It is possible to control the quality of capsule endoscopic images using artificial intelligence, but it requires a great deal of time for labeling. Verifying the performance of self-supervised learning for quality control for capsule endoscopic images.

METHODS

A simple framework for contrastive learning of visual representations (SimCLR), is capable of acquiring the inherent image representation with minimal annotation, but the feasibility is not studied. A total of 62 840 images were collected to train models in internal cross-validation (more training data and less testing data) and reversed cross-validation (less training data and more testing data). Random forest and eXtreme Gradient Boosting (XGBoost) were used to complete the quality control after SimCLR extracted the features from images.

RESULTS

Random forest and XGBoost reported that the mean area under the receiver operating characteristic (AUROC) curve exceeded 0.98 and 0.97 using SimCLR-derived features. Moreover, XGBoost surpassed supervised convolutional neural network (CNN). Extra 12 032 images were gathered for prospective validation and the AUROC of SimCLR surpassed 0.93 (95% confidence interval = 0.9271-0.9548), which is close to supervised CNN (0.9645) in cross-validation. Moreover, the AUROC of random forest and XGBoost (trained with SimCLR-derived features) surpasses 0.96, which is better than supervised CNN (0.8374) in reversed cross validation.

CONCLUSION

Our results demonstrate that the SimCLR framework enables high-performance quality control in capsule endoscopy, achieving accuracy comparable to - or even surpassing - supervised learning methods, while requiring substantially fewer annotated images.