Assessing response in endoscopy images of esophageal cancer treated with total neoadjuvant therapy via hybrid-architecture ensemble deep learning

TL;DR

This paper introduces a deep learning model to accurately assess esophageal cancer patients' response to therapy using endoscopy images, potentially avoiding surgery.

Contribution

The novel contribution is EC-HAENet, a hybrid-architecture deep learning model that outperforms existing methods in predicting treatment response.

Findings

EC-HAENet achieved an AUC of 0.98 in the training cohort and 0.99 in the external evaluation cohort.

The model's accuracy was significantly higher than endoscopic biopsy in both cohorts (0.93 vs. 0.78 and 0.93 vs. 0.71).

Abstract

Esophageal cancer (EC) patients may achieve pathological complete response (pCR) after receiving total neoadjuvant therapy (TNT), which allows them to avoid surgery and preserve organs. We aimed to benchmark the performance of existing artificial intelligence (AI) methods and develop a more accurate model for evaluating EC patients’ response after TNT.

We built the Beijing-EC-TNT dataset, consisting of 7,359 images from 300 EC patients who underwent TNT at Beijing Cancer Hospital. The dataset was divided into Cohort1 (4,561 images, 209 patients) for cross-validation and Cohort 2 (2,798 images, 91 patients) for external evaluation. Patients and endoscopic images were labeled as either pCR or non-pCR based on postoperative pathology results. We systematically evaluated mainstream AI models and proposed EC-HAENet, a hybrid-architecture ensembled deep learning model.

In image-level classification, EC-HAENet achieved an area under the curve of 0.98 in Cohort 1 and 0.99 in Cohort 2. In patient-level classification, the accuracy of EC-HAENet was significantly higher than that of endoscopic biopsy in both Cohorts 1 and 2 (accuracy, 0.93 vs. 0.78, P<0.0001 and 0.93 vs. 0.71, P<0.0001).

EC-HAENet can assist endoscopists in accurately evaluating the response of EC patients after TNT.

Figures