Computer-aided diagnosis using deep learning technology may be a valuable addition to endoscopy surveillance for early cancer in patients with Barrett’s oesophagus (BE).  

European research has led to the development of a computer-aided detection (CAD) system with high accuracy at detecting lesions on endoscopic images.

Importantly, the system works at a speed that can be applied in real time during endoscopy to flag suspicious lesions which endoscopists may then wish to biopsy. 

The deep learning CAD system was “pre-trained” with a data set of almost 500,000 endoscopic images from the oesophagus, stomach, small bowel and colon. It was then further trained and internally validated in a stepwise fashion on more than 1,500 images from two data sets of early BE neoplasia and non-dysplastic BE.

To validate their CAD system, researchers tested it on two separate image datasets including a challenging set of 80 images considered by experts to contain subtle neoplastic lesions.  

The performance of the system in this final set of images was then benchmarked against the performance of 53 endoscopists from four countries. They included novice through to senior level endoscopists, none of whom were considered expert in BE endoscopy. 

The study found the CAD system had a high degree of accuracy – correctly identifying 36 of 40 images of neoplasia and 35 of 40 images of non-dysplastic BE in the fourth data set and 37 of 40 images of dysplasia and 33 of 40 images of non-dysplastic BE in the more challenging fifth data set. 

The performance of the CAD system was also superior to non-expert endoscopists on the basis of accuracy (88% v 73%), sensitivity (93% v 72%) and specificity (83% v 74%). 

The mean time for classification of the image using the CAD system was 0.1 seconds. 

“The high diagnostic accuracy that the system achieved suggests that the CAD system should perform well in a general clinical setting, outside major referral centres that specialise in BE assessment,” the study authors said. 

“We anticipate that the system’s performance will further improve by expanding the number of images in the different database layers of our deep learning system and by including video recordings.”

Machine learning is the future but not quite yet 

The research adds to the growing evidence – for example here and here – for the use of deep learning techniques in the classification of colonic polyps.

Similar work is also underway locally at the University of Adelaide which hosts the Australian Institute for Machine Learning.

Dr Johan Verjans, deputy director of medical machine learning at the Institute, told the limbic artificial intelligence was likely to augment clinical decision-making rather than replace clinicians. 

However there were still a number of challenges before it would be widely available in clinical practice. 

“It’s still relatively far away. It can do very specific tasks really well, but the penetration in the clinic is still pretty rare except for example IVF embryo selection which was developed in Australia.”

“Reproducibility is one big challenge. Is it applicable to different data sets; to different cohorts?”

“AI can be fooled relatively easily so you need to train it on different scanners or cameras. If you train it on a 6-megapixel camera that doesn’t mean it works on a 12-megapixel camera. Same with MRIs,” he said.

“And there is still a lack of RCTs to test the true value of translating AI applications from ‘nice to have’ into really useful applications that outperform, save costs and or do things quicker than doctors.” 

There were also ethical considerations similar to those about who could be held responsible when using self-driving vehicles.