Explaining Deep Survival Analysis Models for Heterogenous Data

Explaining Deep Survival Analysis Models for Heterogenous Data

Title: Explaining Deep Survival Analysis Models for Heterogenous Data


Type: Master Thesis

Student: Moritz Wagner

Supervisor: Sebastian Pölsterl, Christian Wachinger

Status: Running



In my thesis, I aim to predict the progressions from mild cognitive impairment (MCI) to Alzheimer based on 3D MRI brain images and biomarker data. To pursue this, I will leverage state of the art methods in the area of Deep Survival Analysis. Yet, in the medical domain, it is an almost unconditional requirement to not only achieve high prediction performances but to also be able to explain the results. While the interpretation of the biomarker effects is rather straightforward, it is still unclear how to interpret the effects derived from the brain images. The learned latent representation has no substantive meaning so that interpreting its effects is hardly meaningful. Therefore, one aims to directly determine which structures in the brain are responsible for an either, accelerated or decelerated disease progression. To pursue this, it is desired to relate structural differences between an observed unhealthy brain and a hypothetical healthy brain to the differences in prediction outcomes. While methods for assigning feature attribution values are already broadly covered by the literature, it remains a challenge to derive the hypothetical, ‘baseline’ image, as it must meet two criteria. First, it shall depict a realistic image and secondly, it shall only differ in its most salient structures from the unhealthy brain image. By adopting the idea of image-to-image translation from the field of generative modeling, I will propose a method that allows to generate images that fulfill these criteria. To verify the validity of this approach, I will also work with simulated data and 2D slices of the MRI.