AI expert collaborative generative model for predicting pathologic response to neoadjuvant chemoimmunotherapy in NSCLC

Accurate prediction of pathologic response to neoadjuvant chemo-immunotherapy for NSCLC at baseline is important for personalized perioperative treatment selection. However, as a new therapy, there is a lack of sufficient clinical cohorts for model development. Our study aims to develop and validate an AI-expert collaboration generative integration system (AEGIS) to predict the pathologic response to neoadjuvant chemo-immunotherapy for NSCLC using baseline CT. In this retrospective study, a total of 320 patients were recruited from two centers. We collected CT scans taken before neoadjuvant therapy. We established an integrated framework that included a stable diffusion model, LoRA fine-tuning, quality assessment, and reinforcement learning. AEGIS was trained using a mixture of real and synthetic CT images and was validated on an external cohort. This study successfully constructed AEGIS through fine-tuning using a NSCLC immunotherapy dataset and AI-expert collaboration. In the dichotomous model (PR vs Non-PR), AEGIS improved the AUC of the external validation cohort from 0.642 to 0.784 through 500% mixed training. In a trichotomous model (CPR, MPR, Non-PR), AEGIS can increase the accuracy of the external validation cohort from 62.56% to 78.97%. In regression analysis, AEGIS improved the R2 of the external validation cohort from 0.327 to 0.525. Regression analysis outperformed dichotomous analysis in the external cohort (0.752 versus 0.642). This proof-of-concept study shows that AI-expert collaboration generative models can effectively expand sample size and improve model training, thereby achieving more accurate predictions of pathologic response to neoadjuvant therapy for NSCLC. Lung cancer is one of the most common cancers worldwide. For patients with non-small cell lung cancer, doctors sometimes give chemotherapy and immunotherapy before surgery to shrink the tumor. Knowing in advance whether a patient will respond well could help doctors personalize treatment decisions. We developed an AI system that predicts treatment response from CT scan images taken before therapy begins. A major challenge is the limited availability of patient data for this new treatment. To address this, the AI system can create virtual medical data. Using a mix of real and virtual medical data improved the performance of the AI system. This approach offers a promising strategy to build better prediction models where patient data remains scarce.