Longitudinal Radio-Anatomical Modeling and Discovery of Prognostic Factors via Artificial Intelligence: an Ablation Study

Response evaluation is a crucial aspect in the field of oncology as it allows clinicians to assess the effectiveness of anti-cancer treatments, make adjustments to management plans, and determine the overall prognosis of patients. The widely adopted quantitative tool for this purpose is the Response Evaluation Criteria In Solid Tumors (RECIST), which classifies therapy response based on one-dimensional diameter measurements of target lesions, categorizing them as partial response, stable disease, or progressive disease. However, RECIST has certain limitations, including inter- and intra- observer variability, as well as reliance on one-dimensional measurements only. These limitations can impact the accuracy of assessments and subsequently affect patient prognoses. Therefore, there is a need for a new method to overcome these drawbacks. Inspired by the classic radiological reporting approach that identifies all changes throughout the entire body, we can formulate the problem as an image-to-image registration task using neural networks. In this framework, anatomical changes between follow-up scans of the same patient are represented as deformation fields, and these deformations are utilized to predict survival, assuming that they hold valuable prognostic information. While this has been proposed in a few pilot studies, yielding significant results, it remains unclear whether the network's ability to model deformation fields is directly correlated with its ability to predict survival. This thesis aims to address this question through an ablation study, wherein different components of the network architecture are removed or modified to introduce variations in registration quality and examine their impact on survival prediction. The study design includes four experiments, plus an additional one, each analyzing different combinations of network components. These include variations in network size, expressed as features number, inclusion of skip layers, realism of reconstruction implemented via Generative Adversarial Networks (GANs), representation via Vision Transformers, and influence of embedding vectors via latent-space similarity. Survival prediction of the resulting models has been applied to an internal dataset consisting of thoraco-abdominal CT scans from patient who underwent immunotherapy between 01/01/2013 and 31/12/2018 at The Netherlands Cancer Institute-Antoni van Leeuwenhoek Hospital (NKI-AVL; Amsterdam, The Netherlands).