Radiomics using disentangled latent features from deep representation learning in soft-tissue sarcoma

Timothy Sum Hon Mun, Amani Arthur, Imogen Thrussell, Jessica Winfield, Dow-Mu Koh, Paul Huang, Christina Messiou, Simon Doran, Matthew Blackledge

Keywords: MRI, Diffusion-Weighted Imaging, Radiomics, Deep Learning, Representation Learning

Abstract: Detecting the treatment response of radiotherapy for soft-tissue sarcomas (STS) is difficult due to the intratumoral heterogeneity of the disease within tumours. Radiomics and deep learning offer opportunities to find novel biomarkers of treatment response. Small sample sizes are also a common challenge when using modeling problems due to it being rare cancer. We investigate the use of representation learning by developing an unsupervised deep learning pipeline that can learn disentangled and interpretable features from the data. We look at the pairwise correlation between the features and also evaluated their baseline test-retest repeatability because having good baseline repeatability is important for these features to be useful for personalized treatment. We demonstrate that the features generated by this approach using are both pairwise independent and stable upon repeat baseline measurement. We also investigated the use of these features alongside clinical features in predicting recurrence-free survival in a pilot cohort.