[ Paper Summary ] Real-time differentiation of adenomatous and hyperplastic diminutive colorectal polyps during analysis of unaltered videos of standard colonoscopy using a deep learning model
Again, more and more of these papers shows what deep learning can offer to health care industry.
Please note that this post is for my future self to look back and review the materials on this paper without reading it all over again.
Abstract
Computer aided systems can help, in differentiating diminutive adenomas from hyperplastic polyps with high accuracy. And in this paper, the authors have showed an example use-case of where they take a convolutional neural network in order to perform classification on polyps.
What is already known on this subject? / What are the new findings?
Experts have good results in general but community endoscopists fall short of keeping some Preservation and Incorporation of Valuable endoscopic Innovations (PIVI) guidelines. AI may be an assist to those areas. And the authors of this paper have showed a use case where real time classification is possible using deep learning.
Introduction
During colonoscopy, Endoscopists makes difference decisions such as classification of cancer types and regions where to remove or where to keep. And during the procedure, image analysis can aid the Endoscopists. Currently Image analysis, have achieved success in accurately determining, the histology of diminutive lesions which have minimal risk of cancer. (And in turn this have financial benefits.). However these approaches are usually operator depended hence there are some cases where community physicians’s accuracy is below the accepted performance thresholds.
And to overcome these matters, different cost effective methods have been researched. And one of the results of that was the NBI International Colorectal Endoscopic (NICE) classification system. And this system was developed to aid endoscopists who does not have extensive experience. But this method is not able to address problems such as sessile serrated polyps (SSPs).
More recent development of automated image analysis were developed to improve lesion detection. Traditional image analysis techniques depend on hand crafted features, and they are not robust in some cases. But the development of deep learning have the potential to over come this problems. And the authors of this paper used deep convolutional neural network to perform classification of different types of polys.
Methods
Using the NBI videos captured on Olympus colonoscopes, the authors of this paper successfully trained a Deep convolutional neural network. And all of the used data were de-identified before being used in this study.
One advantage of the DCNN was the fact that humans do not have to make hand craft features for the models to successfully perform well. And the general network architecture can be seen above. (They used SGD, mini batch of 128, as well as performed data augmentation of flipping.) Finally there were two classes for the polyps conventional adenoma or a serrated class lesion.
As seen above, one interesting fact about the frame work is the notion of credibility. Where it indicates the confidence level of the model. If the credibility level was below 50 percent, those predictions were excluded from the accuracy calculation. All of the used video were 10–20 seconds with the median of 16 seconds.
Results
125 polyp videos were used as a test set, however among these data only 106 were used to calculate the accuracy of the model since their had high enough credibility levels.
And for the 106 polyp the accuracy was 94 percent, and sensitivity for identification of adenomas was 98%, specificity was 83%, negative predictive value was 97% and positive predictive value was 90%.
Discussion
Colorectal cancer is the third leading cause of cancer death in USA and Colonoscopy plays a critical role in not only identifying but also treatment of this diseases. Although not all of the aspect of Colonoscopy can be replaced by a software, detection of polyps and prediction of histology are the two aspects where software can play a role. And the authors of this paper have showed that it is possible to use an AI to perform these tasks with high accuracy. (Even enough to say that the model have performed better than some community endoscopists.).
Final Words
Although the authors of this paper have discarded few of the testing data while measuring accuracy, this paper delivers promising results. And was very well written.
Reference
- Byrne MF, e. (2018). Real-time differentiation of adenomatous and hyperplastic diminutive colorectal polyps during analysis of unaltered videos of standard colonoscopy … — PubMed — NCBI . Ncbi.nlm.nih.gov. Retrieved 9 July 2018, from https://www.ncbi.nlm.nih.gov/pubmed/29066576
