Federated Voxel Scene Graph for Intracranial Hemorrhage

TL;DR

This work proposes the first application of Federated Scene Graph Generation and shows that the models can leverage the increased training data diversity and can recall up to 20% more clinically relevant relations across datasets compared to models trained on a single centralized dataset.

Abstract

Intracranial Hemorrhage is a potentially lethal condition whose manifestation is vastly diverse and shifts across clinical centers worldwide. Deep-learning-based solutions are starting to model complex relations between brain structures, but still struggle to generalize. While gathering more diverse data is the most natural approach, privacy regulations often limit the sharing of medical data. We propose the first application of Federated Scene Graph Generation. We show that our models can leverage the increased training data diversity. For Scene Graph Generation, they can recall up to 20% more clinically relevant relations across datasets compared to models trained on a single centralized dataset. Learning structured data representation in a federated setting can open the way to the development of new methods that can leverage this finer information to regularize across clients more effectively.

Figures (4)

• Figure 1: Overview of the origin and diversity of the four datasets used for this study: INSTANCE2022, BHSD, CQ500, and a private cohort from Germany. We show the outline of ICH, the ventricle system, and midline. Bleeding 1 from INSTANCE2022, CQ500, and the private cohort all involve the ventricle system, which often serves as a buffer for other brain structures. The ventricle system can compress to absorb external pressure, or conversely it can fill with blood with possible expansion. Such changes are often accompanied by a midline shift, as in the samples of the INSTANCE2022, BHSD and private cohort datasets. Additionally, some images show the results of a previous surgical operation such as the presence of a ventricular drainage (appearing as a white dot within the slice) or even a craniectomy, see the red arrows. \ref{['sec:exp']} offers detailed statistics over these cohorts.
• Figure 2: Examples of the diversity in manifestation of ICH. The outline of the bleeding is shown in yellow. Hemorrhages such as in (a) may require a surgical intervention to repair any ruptured blood vessel or the placement of a drainage to relieve pressure. Similarly, involvement of the ventricular system as in (b) can cause occlusive hydrocephalus and will also require a drainage for the accumulating cerebrospinal fluid. Intraparenchymal (c) and epidural (d) hemorrhages, while dissimilar in appearance, can both cause midline shifts (c and d). Such a shift is associated with increased intracranial pressure and may require surgery.
• Figure 3: Qualitative results: predicted segmentation and relations for a patient from the BHSD dataset by MOTIF trained on the INST dataset (left) and using Fed-MOTIF (right). Both models detect the intraventricular bleedings 1 and 2 and that they are in the ventricle system through the corresponding relation. The model trained on INST already provides a worse segmentation of bleeding 2. It also vastly undersegments bleeding 3 to the point where the same ICH instance also gets detected as an additional bleedings 4. The model trained with FedAvg localizes bleeding 3 much more precisely and detects correctly a ventricle system asymmetry. Additionally, a small subarachnoidal bleeding 4 is only detected by this model (right). Though, both models fail to detect a midline shift.
• Figure 4: Distribution of bleedings and relations for each dataset. All datasets show in bleeding representation whether regarding their number, volume or type. The bleeding types refer to: 1) intraparenchymal, 2) epidural or subdural, 3) intraventricular, 4) basal subarachnoidal, and 5) non-basal subarachnoidal. "Basal" refers to the basal cistern, where the subarachnoidal bleeding can be more prominent.