Probability Level Fusion for Ovarian Cancer Diagnosis Using Multimodal Data

Abstract

Artificial Intelligence (AI) has become an important tool in medical diagnostics by enabling models to learn complex relationships between historical patient data and current clinical indicators, supporting earlier and more accurate disease detection. Ovarian cancer remains particularly challenging to diagnose because it is often detected only in advanced stages, significantly reducing the effectiveness of treatment and increasing the risk of mortality. To address this problem, this article proposes a multimodal ensemble diagnostic framework that integrates heterogeneous data sources including tissue (histopathological) imaging, ultrasound imaging, and clinical parameters into a unified predictive system. The framework employs independently trained modality-specific models whose probabilistic outputs are combined using statistically grounded ensemble strategies. Instead of requiring strictly aligned multimodal datasets, the system uses stochastic pairing of modality predictions. The study uses the concept employed in the Naive Bayes Classifier where the classes are assumed to be conditionally independent given the class, and the prediction is made on this assumption. The final model trained on this combination of predictions learns to make predictions in cases of model agreement and disagreement. In the end we are able to show that the model can generate stable results for all cases of patients, and give reasonably good performance.