Weakly-Supervised Semantic Space Structuring: Cardiac Cycle Position for Cerebral Emboli Visualization Using Contrastive Learning

Transcranial Doppler (TCD) is an important tool for monitoring cerebral emboli, a primary cause of ischemic stroke, by detecting high-intensity transient signals (HITS). The link between HITS and their position in the cardiac cycle is emphasized in the literature, but little exploited for microemboli characterization. We therefore propose to integrate this information directly into the structure of latent spaces derived from unsupervised deep learning models toward a more interpretable data visualization. Specifically, we combine autoencoders (AEs) with a contrastive learning framework, where HITS with the same cardiac cycle position are brought closer together, while those with different positions are pushed apart. This weakly-supervised approach facilitates a guided semantic structuring of the latent space, while preserving the underlying data structure. We evaluate our method on a TCD HITS dataset and demonstrate a 11.85% improvement in the local structuring of latent spaces with respect to cardiac cycle position continuity. This trend is further validated in a more complex scenario where the latent space is structured based on both HITS type and cardiac cycle position, outperforming the AE baseline by 5.92% in terms of category continuity in the latent space. These results validate the use of cardiac cycle positioning of HITS for their characterization and demonstrate its benefits for structuring latent representations. The code is available at https://github.com/MathildeDupouy/contrastive-cardiac-cycle/.