SELF-SUPERVISED MULTIMODAL LEARNING FOR EARLY CANCER DETECTION ACROSS IMAGING AND GENOMICS

Abstract

Early cancer detection remains critical for improving patient survival, yet conventional screening methods face limitations in sensitivity and specificity. This research develops a self-supervised multimodal learning framework that integrates medical imaging and genomic data for early cancer detection without requiring extensive labeled datasets. We collected data from 612 patients including individuals with early-stage cancers and high-risk populations undergoing surveillance. The framework employs contrastive learning to align imaging features from CT scans and MRI with genomic signatures from liquid biopsies and tissue samples. Our self-supervised pretraining strategy learns robust representations from 4,850 unlabeled samples before fine-tuning on labeled detection tasks. The model achieved 91% sensitivity and 88% specificity for detecting stage I cancers across multiple tumor types, outperforming supervised baselines by 14 percentage points. Feature analysis revealed that the self-supervised approach discovered clinically meaningful imaging-genomic associations including radiological patterns correlating with specific mutation profiles and circulating tumor DNA levels. The framework demonstrated particular strength in detecting cancers missed by single-modality approaches, identifying 34% more early-stage tumors than imaging alone and 28% more than genomic screening alone. These results demonstrate that self-supervised multimodal learning can overcome labeled data scarcity while improving early cancer detection performance through complementary information fusion across imaging and genomic modalities.