2558 - Clinically Interpretable Integration of Molecular Omics and Medical Imaging for Cancer Classification Using Optimal Transport-Based Mapping

Purpose/Objective(s): Accurate cancer diagnosis in radiation oncology depends on imaging, yet radiologic features alone do not fully reflect the molecular heterogeneity that drives tumor behavior and therapeutic response. Molecular omics data offer complementary biologic information but are rarely integrated with imaging in a clinically interpretable manner. The hypothesis is that integrating spatially mapped molecular omics with imaging improves cancer diagnosis compared with single modality approaches. The purpose of this study was to develop and evaluate an interpretable multimodal deep learning framework that directly links molecular measurements to spatial image features and quantifies diagnostic performance gains.

Materials/Methods: We developed a unified multimodal deep learning framework that transforms high dimensional tabular omics data into two dimensional image-like representations using an Optimal Transport–based mapping strategy. These molecular maps were incorporated as additional channels alongside medical images, allowing joint training within a conventional convolutional neural network. The approach was evaluated on publicly available cancer datasets integrating whole-slide pathology images with spatial transcriptomics, as well as neuroimaging datasets combining MRI with molecular measurements. Model performance was compared against image-only baselines. We further examined class-specific contributions of different omics modalities and analyzed spatial attention patterns to assess whether molecular integration enhanced biological and radiologic interpretability.

Results:

Integrating molecular data with imaging improved cancer classification accuracy by up to 5% compared with image only models, with corresponding gains in macro F1 score. Multimodal models demonstrated more consistent performance across tumor classes. Spatial analyses indicated that the network focused on histopathologic and radiologic regions associated with known biologic features. Modality contribution analysis identified distinct omics types driving predictions for specific cancer subtypes, consistent with prior biologic evidence.

Conclusion:

An interpretable multimodal framework that spatially maps molecular omics data and integrates them with diagnostic imaging improves cancer classification performance and provides insight into biologically relevant image regions. This approach supports clinically meaningful integration of imaging and molecular data for oncology applications.