Radiomics Initialized Deep Embedding Network (RIDE-Net) to Prognosticate Survival in Renal Cancers

Brennan Flannery, Thomas DeSilvio, Satish E Viswanath

Confirmation: I have read and agree with the IEEE BHI 2025 conference submission's policy on behalf of myself and my co-authors.

Keywords: deep learning, radiomics, computed tomography, semi-supervised

TL;DR: Using radiomics as a pre-training target for CNNs significantly improves performance, reduces variability, increases reproducibility, and allows models to maintain performance with only 5% of the training data for kidney cancer survival prediction.

Abstract: Popularly used machine learning approaches for medical imaging offer complementary advantages for disease characterization. Hand-crafted computational features (or radiomic features) offers generalizable quantitative measures which can be intricately linked to disease phenotypes, while iteratively optimized convolutional neural networks (CNNs) offer complex features with robust performance to imaging variations. To address the question of how best to exploit the relative advantages of both approaches, we introduce Radiomics Initialized Deep Embedding Network (RIDE-Net) which leverages intuitive radiomic descriptors to enhance the performance of a CNN model. Our approach involves: (1) identifying disease-specific radiomics features associated with an end-point of interest, (2) pre-training a residual learning network to directly predict these specialized radiomic features, and (3) optimizing this primed RIDE-Net to predict the outcome of interest. We evaluate RIDE-Net in the context of prognosticating overall survival for renal cancers using a multi-institutional cohort of 510 patients and a pre-training cohort of 5195 patients with CT volumes. A RIDE-Net based survival model achieved c-indices of 0.67 and 0.66 in testing and holdout validation, respectively, which significantly outperformed a standard radiomics based Cox regression model (0.56, 0.62) as well as a standard ResNet based survival model (0.60, 0.62) with significantly less variation across training runs. We also found that RIDE-Net deep features achieve increased reproducibility compared to standard radiomic features in terms of intra-class correlation coefficient and can achieve similar prognostic performance even when utilizing 5\% of the training data (c-index of 0.66 in validation via few shot analysis). RIDE-Net thus represents a novel integrated approach to combining the strengths of radiomics and deep learning for robust, accurate, and reproducible predictions in medical imaging tasks.

Track: 3. Imaging Informatics

Registration Id: KRNGTN7Y8DY

Submission Number: 229