Detection of pathologic complete response using deep neural network-based endoscopic evaluation in patients with esophageal cancer receiving neoadjuvant chemotherapy: a nationwide multicenter retrospective study from 46 Japanese esophageal centers.

Background: Detecting pathological complete response (pCR) preoperatively facilitated a non-surgical approach after neoadjuvant chemotherapy (NAC). We previously developed a deep neural network-based endoscopic evaluation to determine pCR preoperatively. Its quality warrants improvement with a larger data series for clinical application.

Methods: This study retrospectively reviewed patients with esophageal squamous cell carcinoma (ESCC) receiving NAC at 46 Japanese esophageal centers certified by the Japan Esophageal Society. Endoscopic images after NAC were collected with clinicopathological factors and long-term outcomes. We randomly selected the same number of patients with Grades 0-1a and Grades 1b-2 based on those with pCR (Grade 3). A deep neural network was used for endoscopic image analyses. A test data set, consisting of 100 photos, was utilized for validation. The sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV) of the deep neural network-based model and experienced physicians were calculated.

Results: The study enrolled 1041 patients, including 354 (33%) patients with pCR, the same number of histological non-responders (Grade 0-1a/1b-2, 352 [33%]/368 [34%]). The median values of sensitivity, specificity, PPV, NPV, and accuracy for pCR detection were 80%, 90%, 89%, 82%, and 85%, respectively. The patients with pCR preoperatively demonstrated significantly better overall survival and recurrence-free survival.

Conclusions: This large-scale study revealed that the deep neural network-based endoscopic evaluation after NAC identified pCR with feasible accuracy. The current artificial intelligence technology may guide an individualized treatment strategy, including a non-surgical approach, in patients with ESCC through prospective studies with careful external validation.