> For the complete documentation index, see [llms.txt](https://rexis.gitbook.io/rexis/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://rexis.gitbook.io/rexis/medical-imaging.md).

# Medical Imaging

### Medical Imaging

Medical imaging encompasses a diverse range of techniques that generate 2D representations of internal anatomical structures. These images are essential for:

* Disease diagnosis
* Monitoring and progression tracking
* Treatment planning

Rexis enables secure AI development on imaging data by integrating **Equivariant Encryption (EE)** into the training and evaluation pipeline.

![](https://lh7-rt.googleusercontent.com/slidesz/AGV_vUcx8UVNymcX94EivTJEynKdhHGI6DAxYji6A7Xr06A--ZFkjRqDpRhCpMV8gi3O0pLr39teZLqTcbm4ATir4EkP2Cs8LyYrBLUp2ZpWLK-wiaICGz6xYpTJxF9CghjT9aD54sxg=s2048?key=_t06K9zDHrDgeHDIPqC4uaCR)

***

#### Radiography (X-rays)

X-rays are one of the most widely used imaging modalities, especially for assessing bones, fractures, and dense tissues. They produce 2D projectional images using electromagnetic radiation.

**Data Support**

X-ray images are typically stored in **DICOM** format, which includes both pixel data and metadata.

Using **EE**, Rexis enables encrypted storage and processing of chest X-rays, allowing models to be **trained and evaluated without accessing raw images**, preserving patient privacy.

**Model Support**

We support secure fine-tuning and inference for X-ray models using EE.

* [**BiomedCLIP**](https://huggingface.co/microsoft/BiomedCLIP-PubMedBERT_256-vit_base_patch16_224)\
  A vision-language model trained on biomedical literature and images; bridges textual and visual representations.
* [**CheXagent**](https://stanford-aimi.github.io/chexagent.html)\
  A model that performs automated radiology reporting and disease localization on chest X-rays.
* [**CheXzero**](https://github.com/rajpurkarlab/CheXzero)\
  Uses self-supervised learning to detect chest pathologies without requiring extensive annotation.

***

#### Pathology

Pathology images are ultra-high-resolution scans (WSIs) of stained tissue samples used to diagnose disease at the cellular level.

**Data Support**

Whole Slide Images (WSIs) are often stored in formats like **SVS** or **TIFF**. These gigapixel images are sensitive and often contain identifying histopathological features.

Rexis supports EE-protected training and evaluation pipelines to ensure privacy **even during high-resolution model development**.

**Model Support**

We support secure fine-tuning of digital pathology models under encrypted settings.

* [**UNI**](https://www.nature.com/articles/s41591-024-03141-0)\
  A pathology foundation model optimized for panoptic segmentation across large histology datasets.
* [**CONCH**](https://huggingface.co/MahmoodLab/conchv1_5)\
  A visual-language model that combines contrastive learning with histopathology image embeddings.
* [**Prov-GigaPath**](https://github.com/prov-gigapath/prov-gigapath)\
  A whole-slide model incorporating provenance to enhance interpretability and auditability in pathology AI.

***

#### Ophthalmology

Ophthalmology relies on imaging of the eye’s internal structures, particularly:

* **Fundus Photography (CFP)**: color images of the retina
* **Optical Coherence Tomography (OCT)**: high-resolution cross-sections of retinal layers

**Data Support**

Retinal imaging data is typically stored as **JPEG**, **PNG**, or **TIFF**. EE ensures that models trained on retinal data **never access or leak raw patient images**, allowing safe collaboration across eye hospitals and research labs.

**Model Support**

We support secure fine-tuning and encrypted inference of ophthalmology models.

* [**RETFound**](https://github.com/rmaphoh/RETFound_MAE)\
  A foundation model for disease detection from retinal images, trained for generalization across imaging domains.
* [**REMEDIS**](https://arxiv.org/abs/2205.09723)\
  A unified learning framework aimed at improving robustness and label-efficiency for medical imaging AI.

***

#### Dermatology

Dermatology imaging includes clinical photos and **dermoscopy**, where magnified skin images reveal fine-grained details used to detect malignancies and other conditions.

**Data Support**

Skin images are typically stored in **JPEG**, **PNG**, or **DICOM**. EE allows dermatological datasets to be stored and processed securely—enabling privacy-preserving AI development even when using personal or sensitive skin imagery.

**Model Support**

We support encrypted fine-tuning and secure inference for skin image models.

* [**Derm Foundation**](https://huggingface.co/google/derm-foundation)\
  A dermatology foundation model providing strong visual embeddings. These embeddings reduce the need for large labeled datasets, improving accessibility and training efficiency for skin disease models.

***