AI Model Matches Endoscopists in Distinguishing Viral Esophagitis Types

When a severely immunocompromised patient develops a sore throat and difficulty swallowing, an endoscopist threading a camera down the esophagus faces an immediate diagnostic puzzle. Two viruses—cytomegalovirus and herpes simplex virus—cause nearly identical-looking damage to the esophageal lining. The endoscopist sees the inflammation, the ulceration, the tissue injury. But which virus is responsible? The visual clues overlap so thoroughly that real-time diagnosis is often impossible. Confirmation requires a biopsy, immunohistochemistry analysis, and a wait of days or longer. Meanwhile, the patient remains untreated, and the clinician must guess which antiviral drug to start.

Researchers at a tertiary referral center set out to solve this problem using artificial intelligence. They built a deep learning model trained on endoscopic images from patients whose viral infections had been confirmed by biopsy—the gold standard. But they did not simply feed the algorithm thousands of images and let it learn. Instead, they employed a strategy called curriculum learning, which mirrors how human clinicians actually acquire expertise. The model began with the clearest, most visually distinct cases, then progressively encountered more ambiguous presentations. This sequential approach, moving from easy to difficult, proved more effective than conventional training methods that randomize case order.

The trained model achieved an area under the receiver operating characteristic curve of 0.796, a measure of diagnostic discrimination. Its balanced accuracy—the average of its ability to correctly identify each virus type—reached 0.680. To benchmark this performance, the researchers asked nine novice endoscopists to examine the same independent test set of images. The endoscopists, working by consensus, achieved a balanced accuracy of 0.710. The machine and the humans were essentially equivalent. The model's performance was not statistically superior, but it was comparable—and critically, it could deliver an answer in real time, during the procedure itself.

The significance lies not in the model outperforming expert clinicians, but in what it enables. An endoscopist using this tool would not need to wait for pathology confirmation to begin treatment. A provisional diagnosis, generated within seconds of image capture, could guide immediate therapeutic decisions. For an immunocompromised patient—someone with advanced HIV, a transplant recipient, or a person undergoing intensive chemotherapy—days matter. Delayed treatment allows the virus to spread deeper into the esophagus, causing more severe ulceration, bleeding, and pain. Earlier intervention, even if provisional, can prevent progression and reduce morbidity.

The curriculum learning strategy proved particularly valuable in handling the visually ambiguous cases that stymie human observers. These are the presentations where the endoscopic features genuinely do overlap, where even experienced clinicians hesitate. By training the model to progressively master these difficult distinctions, the researchers created a tool that performs well precisely where human judgment falters most. The model does not replace the endoscopist or the pathologist. Rather, it sits between the procedure and the definitive answer, offering guidance when time is critical.

The work was conducted at a single tertiary center using biopsy-proven cases, which means the training data came from a population already selected for diagnostic uncertainty—patients whose cases were complex enough to warrant tissue confirmation. This is both a strength and a limitation. The model learned from genuinely difficult cases, not from easy ones. But the findings may not generalize perfectly to other institutions, other patient populations, or endoscopists with different levels of expertise. The next phase would involve prospective validation in real clinical settings, testing whether the model's performance holds when deployed in actual practice, with real patients, in real time.