Unsupervised Skull Segmentation via Contrastive MR-to-CT Modality Translation

TL;DR

This work proposes a fully unsupervised approach to skull segmentation, where the participants do not perform the segmentation directly on MR images, but rather perform a synthetic CT data generation via MR-to-CT translation and perform the segmentation there.

Abstract

The skull segmentation from CT scans can be seen as an already solved problem. However, in MR this task has a significantly greater complexity due to the presence of soft tissues rather than bones. Capturing the bone structures from MR images of the head, where the main visualization objective is the brain, is very demanding. The attempts that make use of skull stripping seem to not be well suited for this task and fail to work in many cases. On the other hand, supervised approaches require costly and time-consuming skull annotations. To overcome the difficulties we propose a fully unsupervised approach, where we do not perform the segmentation directly on MR images, but we rather perform a synthetic CT data generation via MR-to-CT translation and perform the segmentation there. We address many issues associated with unsupervised skull segmentation including the unpaired nature of MR and CT datasets (contrastive learning), low resolution and poor quality (super-resolution), and generalization capabilities. The research has a significant value for downstream tasks requiring skull segmentation from MR volumes such as craniectomy or surgery planning and can be seen as an important step towards the utilization of synthetic data in medical imaging.

Figures (5)

• Figure 1: Inference with the use of the proposed solution. Firstly, the generator from the CUT framework is applied. This is followed by a super-resolution module. Next, we perform histogram matching using a different CT sample from the dataset, which does not need to be correlated with the input MR image. After this, Hounsfield thresholding is applied, followed by binary operations, to achieve the final skull segmentation results. The entire process leverages synthetic CT for skull segmentation via MR-to-CT modality translation.
• Figure 2: The overview of the pipeline for the Contrastive Unpaired Translation and Laplacian Pyramid Super-Resolution Network. Note that we work on 3-D tensors, the 2-D representations are used only for visualization simplicity.
• Figure 3: Sampling procedure for patchwise contrastive estimation of real MR $\leftrightarrow$ synthetic CT, and real CT $\leftrightarrow$ identity CT. We show a 2-D view for better visualization.
• Figure 4: Results of skull segmentation from MR images with the use of MedSAM: (\ref{['bb']}) bounding box and (\ref{['pp']}) point prompt approaches. Bounding box approach failure stems from a fact that segmented skull structure is not propagated through the whole image, and only small parts are captured. For point prompt, the model is unable to identify the skull and propagates segmentation into brain.
• Figure 5: Top: Results of translation and segmentation on defected skulls (from left to right: input MR, matched CT mask, synthetic CT mask, synthetic CT mask with removed implant area). Bottom left: Super-resolution of synthetic skull. Bottom right: MR-to-CT translation on child's skull.