Machine vs Man for Cancer Detection in Barrett's Esophagus

A deep-learning computer-aided detection (CAD) system identified neoplasia with high accuracy and near-perfect delineation performance in Barrett's esophagus (BE), achieving greater accuracy in primary detection than non-specialized endoscopists, an international study found.

That may open the door to the early detection of high-risk lesions without resorting to biopsy, said Jacques Bergman MD, PhD, of Amsterdam University Medical Center in the Netherlands, and colleagues.

The study, online in Gastroenterology, found that in one data set of 80 patients, machine learning was able to classify images as containing either neoplasms or non-dysplastic BE with 89% accuracy, 90% sensitivity, and 88% specificity.

In another 80-patient dataset, the system was found to outperform the classifications by 53 general endoscopists as follows:

• 88% vs 73% accuracy

• 93% vs 72% sensitivity

• 83% vs 74% specificity

Additionally, the system identified the optimal site for biopsy of detected lesions in 97% and 92% of cases, respectively.

"The system detected neoplasia with high accuracy and near-perfect localization," Bergman's group wrote. "The CAD system achieved higher accuracy than a panel of non-expert endoscopist assessors, strongly suggesting that CAD may improve the accuracy of surveillance-detected early BE neoplasia by general endoscopists."

While most surveillance-detected esophageal adenocarcinomas in BE patients can be cured with endoscopic surgery, early neoplasia frequently goes undetected since the bulk of surveillance is carried out by non-expert general endoscopists. "This problem arises because general endoscopists are infrequently confronted with early neoplasia in BE, since progression to neoplasm is rare (<1% per patient year)," Bergman and co-authors wrote. "General endoscopists thus have limited familiarity with the endoscopic appearances of early BE neoplasia."

While web-based interactive training programs are being developed to improve the detection rates of general endoscopists, another option is to enlist the aid of artificial intelligence, with its ability to recognize patterns across large amounts of imaging data.

Study Details

Bergman's group developed its hybrid ResNet-UNet model system using five independent endoscopy datasets. Machine pre-training used 494,364 labeled endoscopic images collected from all intestinal segments and after that, the researchers used 1,704 unique esophageal high-resolution images of rigorously confirmed early-stage neoplasia from patients with BE as well as non-dysplastic BE from a sample of 669 patients.

System performance was assessed using datasets 4 and 5.

Dataset 5 was scored by 53 general endoscopists with a wide range of experience from four countries to benchmark the system's performance. Coupled with the histopathology findings, the scoring of images containing early-stage neoplasia in datasets 2 through 5 were delineated in detail for neoplasm position and extent by multiple endoscopists expert in BE, whose evaluations served as the standard of reference.

Asked for his perspective, Satish Nagula, MD, of Icahn School of Medicine at Mount Sinai in New York City, who was not involved with the study, told MedPage Today that about 5% to 10% of patients with longstanding acid reflux develop BE, and approximately 2% to 5% of these will ultimately develop esophageal cancer.

"Our current strategies for BE involve regular surveillance, usually a routine upper endoscopy every 3 to 5 years in patients with a known diagnosis of BE. The challenge in BE is identifying these precancerous areas during endoscopy," said Nagula.

"The dream was that endoscopic technology would evolve to a point where we could identify areas of dysplasia without taking a biopsy -- that we would be able to identify dysplasia just by the visual appearance during endoscopy alone," he continued.

And while the current generation of scopes has sufficiently high resolution to make this feasible, the challenge is for all endoscopists to be able to accurately identify the minute changes that spell dysplasia, he added.

Nagula predicted that CAD systems such as the one described by Bergman's group are undoubtedly coming to endoscopy in the near future: "Many companies are looking at image systems that can help with the detection of pre-cancerous colon polyps, and this technology would greatly increase the ability of general gastroenterology practitioners to identify the subtle pre-cancerous changes in BE. That would allow patients to get referred to expert centers for endoscopic treatment, which can successfully eradicate these pre-cancerous, or even early-stage malignant lesions."

Improving adenoma detection in BE is a mounting concern, with a recent article calling for efforts to achieve a detection rate similar to that in colon cancer screening.

The authors noted several limitations to the study's generalizability, including the high-quality, expertly collected images in some of the datasets, which may have resulted in selection bias. Furthermore, the CAD system was specifically trained and validated only on Fujifilm Eluxeo imagery. In addition, the CAD system, specifically designed to detect early BE neoplasia in overview with white-light endoscopy, relies on still images, whereas video-based systems are now considered the state-of-the-art approach.

Comment