Deep Learning Predicts Imminent Tumor Progression in Advanced Pancreatic Adenocarcinoma Using Serial CT Scans During Chemotherapy

Advanced pancreatic ductal adenocarcinoma (PDAC) often progresses rapidly during chemotherapy despite initial assessments of stable disease or partial response by Response Evaluation Criteria in Solid Tumors (RECIST 1.1), underscoring the limitations of the current methods for predicting short-term progressive disease (PD). To address this, the study developed a spatiotemporal deep learning framework that integrates convolutional and long short-term memory (LSTM) neural networks to dynamically predict PD at the next follow-up visit using serial computed tomography (CT) scans and baseline clinical variables. The model was trained on a retrospective cohort of 243 patients (415 predicted events, defined as temporal sequences for the next follow-up PD prediction) and evaluated across internal, external, and prospective cohorts. The model achieved area under the curve (AUC) values of 0.77, 0.76, and 0.74, respectively. Performance remained robust across chemotherapy regimens (AG or Gemcitabine-based, FOLFIRINOX, and SOXIRI; AUC 0.68-0.79), PD subtypes (target lesion growth vs. new metastases; AUC 0.72 vs. 0.77), and baseline disease stages (locally advanced vs. metastatic; AUC 0.85 vs. 0.71). This framework enables the noninvasive, real-time prediction of imminent PD in advanced PDAC, facilitating timely treatment modification. Its validated generalizability and reliance on routine clinical data underscore its potential for seamless integration into chemotherapy management.