# Sebastian "Sebo" Diaz

> MIT PhD candidate working on machine learning, medical imaging, domain generalization, and representation learning.

Contact: [sebodiaz@csail.mit.edu](mailto:sebodiaz@csail.mit.edu) · [sdd@mit.edu](mailto:sdd@mit.edu) · [seboddiaz@gmail.com](mailto:seboddiaz@gmail.com)

- [Email](mailto:sebodiaz@csail.mit.edu)
- [Twitter](https://www.twitter.com/DiazSebo)
- [GitHub](https://github.com/sebodiaz)
- [LinkedIn](https://www.linkedin.com/in/sebastian-d-diaz/)
- [CV](https://sebo.sh/cv/)

## Research interests

- Incentivizing generalization beyond the training distribution.
- Exploring synthetic data for improved or otherwise unattainable representations.
- Representation learning and optimal use of embeddings.
- Exploiting numerical structure to enhance algorithms.

## News

- Dec 2025: Invited MIT IMES Seminar talk.
- Sep 2025: Invited MIT IMES Retreat talk.
- Jun 2025: MICCAI paper accepted.
- Apr 2024: Awarded NSF GRFP.

## Publications

### Why Invariance is Not Enough for Biomedical Domain Generalization and How to Fix It

Sebo Diaz, Polina Golland, Elfar Adalsteinsson, Neel Dey · In Review

Invariance is a common inductive bias for domain generalization. However, we show that it can lead to sub-optimal performance. We propose a simple and theoretically grounded intervention, named DropGen, that exploits invariance and domain-specific information achieving high-sample efficiency and competitive performance.

- [Project Page](https://sebo.sh/projects/DropGen/)
- [arXiv](https://arxiv.org/abs/2604.02564)
- [GitHub](https://github.com/sebodiaz/DropGen)

### Robust Fetal Pose Estimation across Gestational Ages via Cross-Population Augmentation

Sebo Diaz, Benjamin Billot, Neel Dey, Molin Zhang, Esra Abaci-Turk, Ellen Grant, Polina Golland, Elfar Adalsteinsson · MICCAI 2025

Current fetal pose estimation methods fail to generalize across gestational ages (GA). We propose a novel method to capture lower GA subjects using exclusively higher GA subjects. We achieve state-of-the-art performance on a challenging clinical dataset enabling more accurate motion estimation.

- [MICCAI Publication](https://papers.miccai.org/miccai-2025/0789-Paper3296.html)
- [arXiv](https://arxiv.org/abs/2509.12062v1)
- [GitHub](https://github.com/sebodiaz/cross-population-pose)

### Design of Novel RF Pulse for Fetal MRI Refocusing Trains using Rank Factorization (SLfRank) to Reduce SAR and Improve Image Acquisition Efficiency

Sebo Diaz, Yamin Arefeen, Borjan Gagoski, Ellen Grant, Elfar Adalsteinsson · ISMRM 2022

Relaxed optimization problem enables significant image acceleration. We show 22% acceleration compared to industry standard with no tradeoffs. Applications in fetal, but pertains to any turbo spin echo sequence.

- [ISMRM Publication](https://archive.ismrm.org/2022/2926.html)

### Fetuses Made Simple: Modeling and Tracking of Fetal Shape and Pose

Yingcheng Liu, Peiqi Wang, Sebo Diaz, Benjamin Billot, Esra Abaci Turk, Ellen Grant, Polina Golland · MICCAI 2025

Analyzing fetal body motion and shape is paramount in pre-natal diagnostics and monitoring. Existing methods mainly rely on keypoints or volumetric segmentations of the fetal body. Keypoints oversimplify the body structure, while segmentation lacks temporal correspondence. To address these shortcomings, we construct a 3D articulated statistical fetal body model based on the Skinned Multi-Person Linear Model (SMPL).

- [MICCAI Publication](https://papers.miccai.org/miccai-2025/0333-Paper4748.html)
- [arXiv](https://arxiv.org/abs/2506.17858)
- [GitHub](https://github.com/MedicalVisionGroup/fetal-smpl)

### Stochastic-offset-strategy enhanced RF pulse optimization with auto-differentiation

Molin Zhang, Nicholas Arango, Sebo Diaz, Jacob White, Elfar Adalsteinsson · ISMRM 2024

Voxel-wise objective function with auto-differentiation for RF pulse optimization has become prevalent. While benefit from the spatial flexibility of desired pattern, conventional fixed-point representation of a matrix fed into the voxel-wise objective function leads to sub-optimal and undesired resultant profile at courser resolution. We assign random spatial offsets to each point centered at the voxel and show superior performance to fixed-point representations.

- [ISMRM Publication](https://submissions.mirasmart.com/ISMRM2024/Itinerary/PresentationDetail.aspx?evdid=5606)