1291 Multi-modal deep learning integrating radiology and pathology images to predict cancer immunotherapy response: a retrospective multi-cohort study

• DOI: 10.1136/jitc-2023-sitc2023.1291
• Publication Date: 2023-10-31T14:38:28Z
• Authors: Yuming Jiang; Zhe Li; Ruijiang Li
• Keywords: Digital Pathology

<h3>Background</h3> There is a critical unmet need for predictive biomarkers of cancer immunotherapy. The tumor microenvironment (TME) plays an important role in determining immunotherapy response and outcomes. Here, we aimed to develop validate multi-modal deep learning model that integrates routine histopathology radiology images predict TME status gastric patients. <h3>Methods</h3> In this retrospective multi-cohort study, developed multitask the simultaneous prediction disease-free survival using CT whole slide haematoxylin eosin (H&amp;E)-stained images. Four classes were defined according five immune assessed by immunohistochemistry. We trained convolutional neural network training cohort 320 patients tested internal external validation cohorts. To fuse H&amp;E images, first concatenate density map different types as global feature whole-slide image CT, resulting multi-channel image. Then feed into classification testing. compared performance with other models (single modal models, single task models). Further, evaluated model's association prognosis ability response. Multivariable analysis was performed logistic regression method see how parameters (gender, age, smoking status, model) affected <h3>Results</h3> deep-learning achieved high accuracy assessing both cohorts (AUC 0.81–0.85). multi-task superior significantly associated overall all (p&lt;0.0001). remained independent prognostic factor adjusting clinicopathological variables including size, stage, differentiation, Lauren histology. Moreover, higher 0.78) than programmed death-ligand 1 combined positive score (CPS), their combination further improved accuracy. <h3>Conclusions</h3> could allow accurate evaluation from Furthermore, predicted outcomes cancer. Exploration data may be promising avenue precision <h3>Ethics Approval</h3> Ethical approval obtained institutional review boards participating centers, patient consent waived analysis. <h3>Consent</h3>