Breast Cancer Detection with Missing Modality

Cancer is one of the leading causes of death worldwide. Among many types of cancer, breast cancer remains one of the leading causes of cancer-related deaths worldwide. Early diagnosis is crucial in increasing survival rates. Although achieving rapid diagnosis is desirable, many limitations, such as data availability, complex mammograms, low resources, and an increase in workload on healthcare professionals, may hinder this. The goal of this study is to develop a robust AI-powered breast cancer classification system that improves accuracy, reliability, and accessibility. Multimodal models were trained using various techniques, with transfer learning applied to both mammograms and clinical data as the baseline. Additionally, four ablation scenarios were tested to replicate missing data conditions: ablating images (clinical-only), ablating text (image-only), random modality drop (50%), and substitution of real images with synthetic GAN-generated data. Each experiment determined whether the model's diagnosing ability improved or decreased under restricted input conditions. Results showed that the clinical-only model yielded the highest F1-score (0.962), followed by random modality drop and GAN-generated inputs with 0.93 and 0.96 F1-scores, respectively. Baseline multimodal model with transfer learning achieved an accuracy of 0.8148. Results show that clinical features alone are strong predictors, and synthetic data can improve robustness in low-resource scenarios.