Automated tumor regression grade assessment and survival prediction in esophageal cancer via weakly supervised multiple instance learning

TL;DR

This paper introduces an AI framework for automatically assessing tumor regression in esophageal cancer, improving accuracy and survival prediction.

Contribution

A weakly supervised multiple instance learning framework for automated tumor regression grade assessment in digital pathology.

Findings

The framework achieved 82.7% classification accuracy for tumor regression grade.

AI-derived TRG scores showed better prognostic value than conventional methods.

Strong agreement was found between AI predictions and pathologist assessments.

Abstract

Esophageal cancer remains a major global health burden and is among the leading causes of cancer-related deaths. Accurate evaluation of tumor regression grade (TRG) after neoadjuvant therapy is essential for assessing treatment response and guiding postoperative management. However, conventional TRG assessment relies heavily on subjective histopathological assessments, leading to considerable inter-observer variability and limited reproducibility. We aimed to develop an objective and automated TRG assessment framework using artificial intelligence for digital pathology.

A retrospective analysis was conducted on 157 patients with esophageal cancer and 1,298 hematoxylin and eosin-stained whole-slide images. Three slide-level pathology foundation models and three multiple instance learning methods were evaluated within a patient-level multiple instance learning framework, enabling weakly supervised TRG prediction based solely on patient-level labels.

The proposed framework achieved a classification accuracy of 82.7% and demonstrated strong agreement with manual pathologist grading. Notably, artificial intelligence-derived TRG score provided superior prognostic stratification compared with conventional assessments, showing significant associations with progression-free and overall survival.

This study presents a foundation-model-driven, patient-level multiple instance learning framework for automated evaluation of TRG in esophageal cancer. This approach offers a standardized, reproducible, and clinically interpretable solution that reduces the workload of pathologists and improves prognostic precision. These findings highlight the potential of weakly supervised AI pathology in advancing personalized treatment assessment and decision-making in digital oncology.

Figures