Conditional Convolution of Clinical Data Embeddings for Multimodal Prostate Cancer Classification

Accurate risk prediction for prostate cancer (PCa) in multiparametric magnetic resonance imaging (mpMRI) is essential for non-invasive diagnosis. Current deep learning approaches, such as convolutional neural networks (CNNs), show promise in predicting Gleason scores (GS) directly from imaging data, yet often overlook valuable clinical data in other data modalities. In this study, we introduce EmbedCondConv a novel multimodal GS prediction model that incorporates clinical information by conditioning CNN kernels on principal components derived from patient clinical data. We tested our models on a public dataset of 921 selected MRI scans and corresponding structured data and compared the prediction performance to baseline models. Our results demonstrate that incorporating clinical information into the model improves GS prediction accuracy with AUROC of 0.90 compared to 0.69 using imaging alone. All code used in this study is publicly available at https://github.com/jiayangz/EmbedCondConv.