Spaces:
Sleeping
Apply for a GPU community grant: Academic project
MorphoBuffer is an open-source research platform for extracting, exploring, and comparing biologically meaningful representations from cancer histopathology images using pathology foundation models.
The project addresses a major accessibility barrier in computational pathology: extracting foundation-model embeddings from whole-slide images is computationally demanding, preventing many students, clinicians, and research groups from experimenting with modern pathology AI. MorphoBuffer will provide a public Hugging Face Space where users can analyze example public-domain histology tiles or upload non-identifiable research images, generate foundation-model embeddings, visualize morphological similarity, inspect attention and tissue representations, and export reusable feature tables for downstream research.
The initial implementation is being developed using publicly available cancer histopathology data and pathology foundation-model encoders. The Space will include several research-oriented workflows:
- GPU-accelerated feature extraction from histology tiles.
- Image-to-image morphological similarity search.
- Interactive two-dimensional embedding visualization.
- Comparison of tissue regions across cancer types and morphological states.
- Downloadable embeddings and analysis-ready feature tables.
- Transparent reporting of model provenance, preprocessing, uncertainty, and limitations.
All application code, documentation, example data manifests, preprocessing procedures, and reproducibility materials will be released openly. The interface will be designed for researchers who do not have access to dedicated GPU infrastructure and will include educational examples explaining how pathology foundation-model representations should and should not be interpreted.
GPU access is required for responsive on-demand inference. The current CPU workflow is suitable for small tests but is too slow for an interactive public application. GPU functions will be invoked only when users request embedding extraction, while previously computed results will be cached to minimize resource consumption. An NVIDIA L4, A10G, or an appropriate ZeroGPU allocation would support the initial deployment.
MorphoBuffer is strictly a research and educational resource. It will not provide diagnoses, treatment recommendations, or patient-level clinical predictions. Uploaded images will not be retained, and the interface will instruct users not to submit identifiable or confidential clinical material.
The long-term goal is to establish a community-accessible pathology AI workbench where researchers can compare open models, contribute analysis modules, reproduce published workflows, and evaluate how histological morphology relates to molecular and biological cancer states.