Purpose: The goal of the study is to assess whether machine learning based models applied to baseline CT scans enable the prediction of patient clinical outcomes after chemo embolization with drug eluting beads loaded with irinotecan (DEBIRI).
Material and Methods: Between May 2012 and December 2016 all patients from the FFCD1201 trial referred to DEBIRI were included in this observational study. All patient clinical outcomes were monitored and labeled retrospectively as either good outcome or bad outcome according to 3-month and 6-month evaluations. After segmentation, our population was clustered and, for each cluster, we produced an image representing the mean of this cluster and the inter subject variability (i.e. an atlas). These atlases were constructed using tools of Riemannian geometry and more precisely, flows of diffeomorphisms in the Large Deformation Diffeomorphic Metric Mapping (LDDMM) framework. The model base is shown in figure 1.
Results: Sixty colorectal liver-dominant consecutive patients treated with DEBIRI were included and 292 CT scans were available. Patients were labeled retrospectively as either having a good outcome or bad outcome according to 3-month and 6-month disease status. From those CT scans, two atlases were constructed, one for the 3-months responses scans and another for the 6-months responses labeled as good or bad clinical outcomes. We used leave-one-out cross-validation to evaluate our classification. The mean atlas and both good and bad outcome atlases are shown in figure 2.
Conclusions: Building an atlas for good responders and bad responders is feasible based on the baseline CT scan (phase 1). This model should be externally validated on a prospective cohort in order to estimate its sensibility and specificity (phase 2). At a later stage (phase 3), we will apply this method on a longitudinal dataset. This will allow us to predict the patient clinical outcome based on the evolution of the CT scans.
