Abstract 1454: Development of an AI framework for identifying image based digital biomarkers predictive of immunotherapy response in malignant melanoma.

Key Points

• This research aims to develop an AI framework capable of identifying image-based digital biomarkers that predict immunotherapy response in malignant melanoma.
• Assembled a cohort of 200 melanoma patients with digitized histopathological slides and treatment histories.
• Expert pathologists annotated 138 cases for training the AI model, generating detailed region- and cell-level labels.
• Utilized a larger dataset of 585 cases with over one million annotated cells for model pretraining.
• The framework characterizes tumor microenvironments by analyzing tissue compartments and immune infiltration patterns.
• Achieved high segmentation accuracy (91.9-95.7%) for tumor-related tissue classes.
• Extracted features highlighted spatial and morphological characteristics of tumors, illustrating immune infiltration.
• The AI framework demonstrated robust performance in identifying biologically relevant image features for predicting treatment responses.