discussion_and_conclusion.tex

\vspace{-0.25cm}

\section{Discussion and Conclusion} \label{sec:discussion}
    From a comparison of errors and a visual analysis it appears that the likelihood based models provide the best results most often.
    
    The model used for the DeepHit model had more parameters than the model used for all other methods (due to the output being larger), thus it may not be an entirely fair comparison. However, while using a larger model the method does not provide results significantly better than the likelihood models.
    
    When clinical features were used it seems to improve results, although not significantly. For the increase in complexity it may not be worth including.
    
    What is not factored into the results is computation time. The Cox loss without \gls{MB} is the fastest to compute, the likelihood losses are not much longer. The DeepHit loss takes slightly longer than both previous methods while the Cox \gls{MB} loss takes magnitudes longer (approximately six hours vs four days).

    % In the future it may be interesting to take the survival time output from the models presented here and use them as part of a generative model to predict \gls{CT} acquisitions with a lower survival time.