Artificial intelligence predicts pathological response to neoadjuvant radiochemotherapy from diagnostic biopsies in esophageal adenocarcinoma

Abstract The incidence of esophageal adenocarcinoma is rising globally, and the disease remains difficult to treat. Locally advanced disease requires multimodal therapy. The ESOPEC study identified perioperative chemotherapy as the standard of care. However, randomized trials guide treatment decisions for the majority of patients and are not designed for individualization. Notably, less than 50% of patients respond to FLOT therapy (defined as tumor regression with <10% viable tumor cells post-surgery), suggesting that carefully selected patients may still benefit from CROSS. Artificial intelligence (AI) transforms personalized medicine by enabling various predictions for individual patients, but identifying responders to specific therapies remains a key challenge. This study explores AI for predicting individual responses to preoperative radiochemotherapy following the CROSS protocol for esophageal adenocarcinoma. Our AI model, trained using 312 images from H&E-stained primary tumor biopsies, achieved AUROCs of 0.82 (validation, n = 31) and 0.80 (test, n = 39), predicting 39% and 36% of patients as responders, respectively. Cell nuclei segmentation and interpretability techniques revealed that responders may be distinguished by a higher proportion of tumor cells and spatial tumor cell clustering, along with fewer lymphocytes and less connective tissue. This AI-driven method can potentially inform future treatment decisions for esophageal adenocarcinoma patients by identifying those likely to respond to radiochemotherapy. Such patients could benefit from a personalized treatment strategy featuring radiochemotherapy beyond the standard chemotherapy regimen.