eduzhai > Applied Sciences > Engineering >

Interpretable Deep Models for Cardiac Resynchronisation Therapy Response Prediction

  • Save

... pages left unread,continue reading

Document pages: 10 pages

Abstract: Advances in deep learning (DL) have resulted in impressive accuracy in somemedical image classification tasks, but often deep models lackinterpretability. The ability of these models to explain their decisions isimportant for fostering clinical trust and facilitating clinical translation.Furthermore, for many problems in medicine there is a wealth of existingclinical knowledge to draw upon, which may be useful in generatingexplanations, but it is not obvious how this knowledge can be encoded into DLmodels - most models are learnt either from scratch or using transfer learningfrom a different domain. In this paper we address both of these issues. Wepropose a novel DL framework for image-based classification based on avariational autoencoder (VAE). The framework allows prediction of the output ofinterest from the latent space of the autoencoder, as well as visualisation (inthe image domain) of the effects of crossing the decision boundary, thusenhancing the interpretability of the classifier. Our key contribution is thatthe VAE disentangles the latent space based on `explanations drawn fromexisting clinical knowledge. The framework can predict outputs as well asexplanations for these outputs, and also raises the possibility of discoveringnew biomarkers that are separate (or disentangled) from the existing knowledge.We demonstrate our framework on the problem of predicting response of patientswith cardiomyopathy to cardiac resynchronization therapy (CRT) from cinecardiac magnetic resonance images. The sensitivity and specificity of theproposed model on the task of CRT response prediction are 88.43 and 84.39 respectively, and we showcase the potential of our model in enhancingunderstanding of the factors contributing to CRT response.

Please select stars to rate!

         

0 comments Sign in to leave a comment.

    Data loading, please wait...
×