Oesophageal cancer, illustration
Cross section showing human oesophageal cancer, computer illustration.

A deep learning system can markedly improve the detection of early-stage esophageal cancers during routine endoscopy, according to a randomized, controlled trial (RCT).

The artificial intelligence system nearly doubled the ability of clinicians to identify high-risk esophageal lesions (HrELs)—consisting of esophageal cancer or precancerous lesions— compared with unassisted endoscopy.

This translated into the detection of one extra positive HrEL case per 111 patients screened, the investigators reported in Science Translational Medicine.

“The present study is the first prospective large-scale RCT to validate the effects of a deep learning–based system to improve the performance of endoscopists in detecting cancerous esophageal lesions,” stated Shao-Wei Li, PhD, an assistant professor at Taizhou Hospital of Zhejiang Province in China, and co-workers.

Their findings could help speed diagnosis and treatment for esophageal cancer, one of the ten most common cancers worldwide.

Neoplastic lesions in the esophagus are difficult to identify, particularly in the early stages when they are prone to superficial and subtle morphological changes such as mucosal redness or erosion, uneven surface, and subtle changes in vascular texture.

To aid their identification, Li and team developed a real-time system for the detection of esophageal lesions based on deep convolutional neural networks (CNN).

The ENDOANGEL–esophageal lesion detection system (ELD) was built using datasets that included more than 190,000 esophagogastroscopic images from three clinics in China.

It was then tested in 3,117 patients, aged at least 50 years, who were consecutively recruited from Taizhou Hospital and randomly assigned to receive CNN-assisted endoscopy or a control group of unassisted endoscopy.

The primary endpoint of HrEL detection rate was significantly higher in participants receiving deep learning assistance than in those for whom it was not used, at a corresponding 1.8% versus 0.9%.

The system’s sensitivity, specificity, and accuracy for detecting HrELs were 89.7%, 98.5%, and 98.2%, respectively, and no adverse events occurred.

The researchers noted there were three false negatives with the system, which they stated was because they did not anticipate the diverse clinical scenarios during model training.

“Therefore, as more centers join and the screening samples expand, we will establish a corrective feedback system, allowing the model to continuously learn from a variety of clinical scenarios and identify atypical, hidden, and occult lesions.”

Nonetheless, the investigators maintained: “In conclusion, ENDOANGEL-ELD system was proven effective and safe for assisting endoscopists in diagnosing HrELs with real-time monitoring.”

They added: “Our prospective, randomized, parallel controlled clinical study assessing a deep learning system in endoscopy screening showed its performance in detecting HrELs.

“Such assistance has clinical value and may help improve the screening and detection rate of esophageal cancer and may also improve patient prognosis by promoting early diagnosis and treatment of esophageal cancer.”

Also of Interest